A plain-language record of what's changed in NotesXML with each release — what we fixed, what we improved, and what's new.
Introducing NotesXML Local. A second edition of NotesXML built from the same code: the complete notebook, media, PDF-annotation, calendar and encryption tool set, with the connected parts — on-device AI, dictation and transcription, cloud sync and licensing — removed at build time rather than switched off. No account, no cloud, no subscription, no ads. It uses the same .nxl files as NotesXML, so notebooks move freely between the two. Sold once per platform: Google Play, the Microsoft Store, and for Linux directly from this site. The 5.199 train (5.199.0–5.199.27) is largely the work of separating the two editions cleanly: Local carries no sign-in, billing, AI-engine or sync components, its own teal icon, its own Help, setup, legal and demo content, and third-party notices that list only what it ships. Nothing about the standard NotesXML feature set changed.
Android: audio recording could not access the microphone. On some Samsung devices (Galaxy Tab S9 FE+, Galaxy S24 Ultra on Android 16) starting a recording failed with “Could not access microphone” even though permission was granted. The app’s own pre-recording check was leaving the microphone open when its short test recording ended early, and the fallback that should have taken over never engaged. Both are fixed — owner-verified on both devices.
Recordings can be stopped and continued. Stopping a recording in the Audio Capture window now pauses it: press Record to carry on where you left off, or Save Note to finish, and everything ends up in one recording. Previously pressing Record again after a stop silently started a new recording and the first was lost. A second recording can no longer be started while one is in progress, and Record in Background simply moves the current recording to the background bar. Also on Android: the level meter no longer flickers in a quiet room, and Save Note no longer needs two presses.
Calendars remember a phone view and a desktop view. A Calendar Note, Task List or Event List now keeps two views: on a phone it opens in the 3-Day view, on Windows and Linux in Month, and switching the view on one kind of device no longer changes the other, even for a notebook you use on both. Task List and Event List ranges are now saved with the notebook too. The 3-Month view no longer draws the three months on top of each other.
Android polish. The + (Insert) panel no longer opens the keyboard over the pinned quick actions — it appears only when you tap the search box. A quick double tap on a page in Page view no longer pops the page menu up over the editor.
Also: the bundled third-party notices state the engine versions actually shipped (llama.cpp b10795 on desktop and Android, whisper.cpp v1.8.4) and the full 16-model catalog; a dead Android entry point left over from an earlier sign-in design was removed from both editions.
Android: sync could delete a Notebooks folder that was not empty. With a chosen storage folder, the tidy-up that removes empty folders after a sync relied on an Android delete that removes everything inside a folder, and the check meant to stop it was reading the wrong answer. It now confirms a folder is genuinely empty on both sides before removing anything, never touches your top-level folders, and does not run at all if the sync reported errors. Notebooks held in Google Drive or Dropbox were never affected and can be downloaded again.
Speech and dictation. Dictation now starts straight away: if your speech model was slower than real time it used to open a warning first and wait, and if that dialog missed your click dictation never began at all. Read Aloud no longer skips a sentence in silence when the speech engine fails on one. Continuous recording asks for the Audio Recording Notice properly instead of quietly refusing to start. On Android, imported audio files can be transcribed again, and audio held in another app no longer needs a slow extra copy first.
AI answers are no longer cut short. Custom prompts, Summarize, Outline, Extract Action Items, Action Plan, Brief Extract and Themes could all stop mid-sentence because a safety limit was sized for short note actions. AI Chat and document analysis now measure what a request actually needs and load a larger working memory when required, so asking about a long document no longer means the model sees it in pieces.
Your place on the page is kept. Adding an image to a gallery, deleting a table row, rotating a picture, checking off a list, renaming an attachment or deleting a note no longer scrolls you back to the top. On Windows and Linux, Settings, AI and other dialogs can now be dragged by their title bar so you can move one aside instead of closing it.
Privacy: a background check that contacted a Google address every 30 seconds to estimate connection speed has been removed. NotesXML now works that out from what your device already reports.
Also fixed: task titles are readable in the calendar and task list on every theme; Ctrl+N and the New Page button create a page again; and the desktop packages now carry the third-party licence notices, which cover all 16 AI models.
The AI model catalog grows from 10 models to 16. Six models join the Professional catalog after a full evaluation: Gemma 4 E2B (vision and audio, Lightweight), Mellum2 12B-A2.5B, Devstral Small 2 24B, Muse Glimmer 30B, Laguna XS 2.1 33B-A3B and Nemotron 3.5 Lightning 30B-A3B. The recommended model is unchanged on Android (LFM2.5 VL 1.6B); on desktop it is now Ministral 3 3B. Three Granite 4.2 models were evaluated in the same round and not adopted — each was matched or beaten by a model already in the catalog at the same memory footprint.
Android: image analysis works again. Since 5.198.20, asking the AI about an image on Android silently returned a text-only answer — the picture was sent but never reached the model, following an AI engine update that changed how text is passed alongside an image. OCR, describe and all other vision features are restored. Android also gained the diagnostics that found it: engine errors now reach the app’s log instead of being discarded.
Zero-knowledge sync: large files were unrecoverable. Files over roughly 10 MB encrypted for zero-knowledge cloud sync were written in a format that could not be read back, while the encryption step reported success. The format is corrected. Files synced before this update remain unreadable from the cloud copy and should be restored from the local copy, which was never affected — if you use zero-knowledge sync with large attachments, re-sync after updating.
AI replies no longer come back empty. With some models, the engine could classify an entire reply as internal reasoning; NotesXML discarded that channel and the feature received nothing even though the model had produced a full answer. The reply is now delivered when it is the only text the model produced.
Also fixed: importing a model catalog file on Android (rejected since May as an unsupported format) and switching back to the built-in catalog afterwards; “Save As” no longer warns that a just-saved notebook was modified on disk; and on Android the catalog now reads “7 of 16 models available on Android” with a list of the desktop-only models, rather than a count that did not match the list. The catalog update signing key has been replaced — no action is needed, this build ships the current catalog already.
This release is almost entirely about sync. It is much faster — a first sync of a large collection on Android now takes a couple of minutes rather than twenty — and it is considerably harder to lose anything to it: your local edits are no longer quietly replaced by another device’s copy, calendars, task lists and contacts merge record by record instead of one file overwriting the other, and something you deleted stays deleted. Alongside that, your work is now saved when the app is closed from outside, a long-standing typing freeze on Linux was finally tracked down, and repeating tasks behave themselves.
Preparing a sync used to be the slow part. On Android, checking which files had changed took about a second per file because of how files are located in shared storage, so a first sync of a large collection could run for twenty minutes; the app now checks everything in one pass and that comparison takes a couple of minutes, with progress reported throughout. Looking at what is in your Google Drive asked Drive about each folder one at a time — roughly a hundred round trips — and now fetches the whole list in a couple of requests, on Windows and Linux as well as Android; that step went from about eighty seconds to a few. Two things that made long syncs worse are also fixed: a sync no longer slows to a crawl when the app is in the background or the screen is off, and a sync that is genuinely working is no longer cancelled by the safety watchdog partway through. One unreadable file can no longer stall the whole run either — it is skipped with a note in the log and the rest continues.
When the same file had changed on two devices since the last sync, one copy used to replace the other outright — and because every launch rewrites your calendar file, calendar groups and task lists were the usual casualties, disappearing and returning as “Recovered” placeholders. Two changes close that. First, the comparison now recognises that both sides changed and treats it as a conflict rather than silently overwriting your copy. Second, the calendar and contact files are merged record by record: groups, events, tasks and contacts from both devices are kept, and where the same item was edited in both places the more recent edit wins. Whether a calendar or task list is shown or hidden is now a per-device choice, so hiding one on your phone no longer hides it on your desktop. Calendar and contact updates also arrive at the start of a sync rather than the end, so they land promptly.
Something you deleted on one device could quietly come back from another. For files and folders, a device that had not yet learned of the deletion would re-upload its old copies; each sync now reads the shared record of deletions before deciding what to send, a file that reappears is only accepted if it was genuinely changed after the deletion, and two devices updating that record at once now have their entries combined rather than one replacing the other. For tasks and events, some delete buttons never registered the deletion for syncing at all — every delete path now does. If something has already come back, delete it once more after updating and it will stay gone. Also fixed: files whose names contain characters such as a long dash now upload to Dropbox instead of failing.
Three ways of leaving NotesXML used to skip saving: restarting the app to apply a setting, the operating system closing it during logout or shutdown, and — on Android — the system removing the app from memory after it had been put in the background. In each case up to thirty seconds of work could be lost with no warning. The app now saves everything it can the moment any of these happen. The Exit menu and window close button already saved and asked about unsaved changes; that is unchanged. Two related fixes: a notebook imported from an XML, CSV or HTML file is now correctly treated as new and unsaved rather than remembering the import file as its home, and if calendar, contact or settings data cannot be written to disk you are now told, where the app previously reported success and even synced the stale file to your other devices.
Some Linux users had NotesXML stop accepting keystrokes for no apparent reason: the text cursor vanished, typing went nowhere, and only a stray Enter or Escape brought it back — in one recorded case after nearly six minutes. The cause turned out to be a confirmation dialog belonging to the operating system opening behind the NotesXML window, where it could not be seen but still held the keyboard. Every such dialog is now drawn inside the NotesXML window, where nothing can hide it — that includes replacing all in Find and Replace, deleting a model, resetting prompts, replacing the demo notebook, deactivating a licence and the cloud re-upload confirmation. The Escape key also works again after using the File menu.
If you had a task repeating every Friday and ticked off this Friday, every future Friday was shown as done too. Ticking off a repeating task now completes only that occurrence, and if you complete different occurrences on two devices before they sync, both survive. A repeating task in the Task List note also shows each occurrence with its own date instead of repeating the first one, which had made the list look like the same task duplicated. Separately, tasks could appear two or more times in a list after being edited; existing duplicates now collapse on their own the next time the list is drawn. And the Day view heading reads Schedule rather than the leftover Classes.
A checklist exported to a standalone web page put each item’s text on the line below its checkbox; exported pages now lay checklists out exactly as the app and the PDF do — re-export anything you exported previously to pick up the corrected layout. The once-a-day overdue reminder no longer reappears on every launch after you have snoozed or dismissed it. AI Chat no longer shows an entire reply as a grey code box when a small model wraps its answer in a code fence. Downloadable model catalogues containing reasoning models now import correctly. And the built-in self test had a number of checks that could never run — probing components removed months ago — which now test the real thing and state a reason whenever they skip.
Tasks you did not finish no longer quietly disappear. Overdue tasks are pinned to the top of your task lists, counted in every calendar view, and gathered once a day into a single summary you can act on — and a reminder missed while the app was closed now catches up rather than being lost. Alongside that, calendars open on today’s date again, task and reminder details stay hidden until you unlock the app, and a long run of transcript and media-folder fixes lands for both desktop and Android.
A calendar note used to reopen on whichever date you were last looking at when it was saved, so a notebook set aside for a few weeks came back showing the past. Calendars now open on today’s date in every view. The view you chose — Day, Week, Month and so on — is still remembered. One related refinement: switching views inside a calendar now stays on the date you are looking at instead of jumping back to today, and the Today button is always there when you want to return.
Tasks whose date has passed no longer vanish from view. Task lists in your notes show an Overdue section pinned at the top, in red, with how many days late each task is, whatever time range the list is set to. A repeating task you missed several times appears as a single entry with a count rather than filling the list with copies — its checkbox completes the oldest missed one, and Complete all and Skip missed handle the rest in one go. Your calendar carries the same information: a red counter in the header of all seven views opens the task list, the Day view lists the overdue tasks themselves at the top of the schedule, and the other views mark the days concerned. Reminders were made to match: a task alarm missed because the app was closed now fires the next time you open it that day (events are unchanged — a meeting that has ended stays ended), and once a day an Overdue Tasks summary gathers everything outstanding into one place, with the same complete, skip and move-to-today actions. How far back the app looks, and whether the daily summary appears at all, are yours to set in Settings.
If you protect NotesXML with a PIN or password, the daily overdue summary and task and event reminders no longer appear until after you have unlocked — including in the first moments of startup, before the lock screen itself has drawn. Nothing is lost by being held back: missed task reminders fire once you are in, and the overdue summary follows shortly after.
Three fixes. Events and tasks whose calendar or task list no longer existed — typically after group changes arrived from another device — were hidden from every view even though the data was still in your file; they now appear under an automatically created Recovered group you can rename or delete, and deleting all calendars or task lists on one device now carries over to your others instead of the old groups returning. Text to speech, which had stopped working, works again. And a completed task can be un-ticked again from the task list widget.
A long run of transcript fixes, most of them visible the moment you export. Transcripts no longer show each time mark twice, and paragraphs now break in the same places on every device instead of differing between Android, Windows and Linux — where time marks were also far too small, or showed 0:00 for video. Exported PDF transcripts no longer break in the middle of a sentence, and a stray short paragraph at the end has been tidied away. Reformatting a transcript no longer renders its paragraphs at heading size or invents page markers that were never there, and the notice about timing being removed now appears when timing really was removed, and not when it was kept. AI actions on transcripts are honest about whether timing is included. Transcription itself reports properly too: the progress bar no longer freezes at 85% and looks stalled, cancelling really cancels, and the summary names the model that actually ran.
Adding a large audio or video file now asks whether to embed it in the notebook or keep it in your Media folder, rather than deciding silently — and the size at which it asks depends on the device, so a desktop can embed far more than a phone. Large audio can be saved on Android too, where it was previously refused as desktop-only. Media already living in your NotesXML folder is no longer copied a second time. Transcribing those files was the bigger fix: it failed outright on Android with a permission error, and elsewhere could fail at the last step, skip the progress window, mis-state the time remaining, produce a summary with pieces missing, or lose its time marks and paragraphs. All of that now behaves as it does for embedded media. Also fixed: video notes remember their place on the page, and dropping a file onto the notebook list shows you the result instead of doing nothing visible — or, with nothing open, splitting it in two.
Repeating events and tasks now behave the way a calendar should. Changing or deleting one asks whether you mean just that occurrence, that one and everything after it, or the whole series — and alarms follow the answer. Alongside that: a defect that let the demo notebook accumulate duplicate contacts, tasks and events is fixed and cleans up after itself, and the Event List and Task List views no longer show a day more than the period they name.
Editing or deleting a repeating event or task now asks whether it applies to only that occurrence, that one and all future ones, or the whole series — everywhere you can edit or delete one, including the full calendar view and list notes. Change the time or details of a single instance without touching the rest, or split a series from a chosen date onwards. Calendar views show each occurrence with its own changes, and ticking off one occurrence of a repeating task no longer marks the whole series complete. Alarms follow your choice: a repeating event now alarms on every occurrence rather than only its first date, honours a time changed on a single occurrence, and stays silent for occurrences you deleted or completed. Editing a series also no longer quietly brings back occurrences you had previously deleted.
Deleting a single occurrence, or ending a series from a chosen date onwards, was recorded correctly but the calendar never acted on it — the occurrences carried on appearing, with no error to explain why. Those markers are now honoured everywhere, in every view and in alarms. Please note: occurrences you deleted in an earlier version may disappear when you upgrade. That is the deletion you asked for finally taking effect, not data being lost.
Loading the demo notebook, or completing the setup wizard, could quietly add another copy of the demo contacts, tasks and event every time — and if cloud sync was switched on, those copies spread to your other devices. This is fixed, and any duplicates already created are cleaned up automatically the next time the app starts, keeping one copy of each. Contacts, tasks and events you created yourself were never affected. A related fault that made contacts vanish from view after a sync until the next restart is fixed too.
Audio recording is now a single button. Opening Audio Capture presents both choices immediately: record here, or record in the background so you can carry on taking notes, drawing or reading PDFs while it runs — the recording continues in a compact bar at the bottom and expands again with one tap. Background recording is a Professional feature; on the free tier the button explains what it does and offers the upgrade. Existing pins for the old Continuous Recording entry carry over, the Help Guide has been updated to match, and saving an audio note no longer writes a harmless internal error into the diagnostic log.
On computers where a large AI model runs on the processor rather than the graphics card, reading a long message before replying can take many minutes — previously with no warning and no way out. AI Chat now learns how fast each model reads on your hardware and tells you up front, in minutes, when a reply is going to take a while, with the option to cancel.
With the Select tool active you can now copy, cut, paste and duplicate any drawing objects — pen strokes, shapes, text boxes and stamps — with the usual keyboard shortcuts, select everything at once, and nudge a selection with the arrow keys. Closing a drawing with unsaved changes now asks first instead of discarding them.
NotesXML now offers six downloadable reading voices, every one trained by Bryce Beattie from public-domain recordings and released to the public domain, so their origins are clear end to end. Speech speed returns to a 1.0x default. Stopping a spoken reply now reliably silences the rest of it, voice mode stops listening when you close the camera, and a rare loss of a spoken sentence during Read Aloud is fixed. Ask Notebook now handles notebooks larger than the AI’s reading space, and the AI engine no longer reports false start failures on processor-only models.
A run of releases centred on PDF annotation. Marks made on a PDF held inside a notebook are now stored where the notebook file actually saves them, so they survive closing and reopening the viewer; highlights drawn over selected text can be erased and selected like every other mark; and the annotation toolbar no longer grows over the viewer’s Close button. On phones and narrow windows, the stamp type and the stroke and fill colours for shapes are reachable from the annotation menu for the first time. The demo notebook gained a Dictate entry and a full table of contents.
Annotations made on a PDF that had been loaded from a notebook file — including the demo notebook — were silently discarded when the PDF viewer was closed and reopened. The annotated copy was being written to a field that the notebook does not save, so every reopen showed the original document again, with no error at any point. The annotated copy is now written to the field the notebook actually stores, so marks survive reopening the PDF and are kept when the notebook is saved. If storing ever fails, the app now says so instead of reporting success.
Highlights made by selecting text behaved differently from every other mark: after closing and reopening the PDF, the eraser and the select tool passed straight over them, and only freehand marks could be removed. Because undo does not survive closing the document, a text highlight became permanent. Text highlights now respond to the eraser and to selection like any other annotation, and one click removes the whole highlight. A highlight sitting on top of another mark no longer prevents that mark from being erased either.
On desktop, switching to Annotate mode grew the vertical tool column over the top bar of the PDF viewer, hiding the Close button and the document title. Leaving the viewer meant switching back to Read mode first. The tool column now stays below the top bar at every window size, and scrolls internally when there are more tools than height.
On phones and in narrow windows, the compact annotation toolbar offered the stamp and shape tools but not the menus attached to them. Whichever stamp was selected by default was the only one that could be placed, and lines, arrows and shapes were stuck in their default colours. The annotation menu now carries a stamp type chooser, and the Colors & Width panel gained stroke and fill colours for shapes.
The built-in demo notebook gained a Dictate a Note entry showing how to speak a note instead of typing it, and explaining how dictation differs from transcribing a recording. Its Welcome page now uses the full Table of Contents note, listing every page and the notes inside it, rather than the older one-level page list.
A run of releases focused on spoken conversation, dictation reliability and AI output quality. Conversation mode in AI Chat now recognises when you have finished speaking — even in a noisy room — and replies in about 1.8 seconds instead of waiting out a timer; it also learns what your room sounds like while nobody is talking. Dictation no longer drops words when it pauses between sentences, and switching the speech model takes effect immediately. AI summaries and transcriptions can no longer run away into repeated text, and anything cut short now says so. The Table of Contents was rebuilt as a live, clickable note with AI page descriptions, the model catalogue gained an in-app update check, and scroll bars are readable in every theme.
In a room loud enough that your voice barely rises above the background, conversation mode could fail to notice you had spoken at all, and every turn ran until the 30-second safety limit before replying. The app was listening for your voice by loudness alone, and in that room your voice never got loud enough relative to the noise. It now also accepts the speech recogniser’s own evidence: if it has transcribed actual words and then stops finding new ones, your turn ends normally. Silence and background noise still never end a turn on their own.
The AI sometimes began a reply with a stray pair of dollar signs — the symbols that normally mark the start of a mathematical formula — with no matching pair to close it. They appeared in the chat text, and in spoken conversation the voice read them aloud. Unmatched formula markers are now removed before a reply is shown or spoken, and the speech side strips them from each piece of a reply as it arrives, so the same thing cannot happen with a genuine formula that arrives in parts. Real formulas still display normally, and prices such as $5 are untouched.
After the last update, conversation mode usually replied about 1.8 seconds after you stopped speaking, but sometimes took 2.7 seconds for no apparent reason. The check that decides you have finished was set to exactly the same moment the app re-reads what it heard, so a difference of one thousandth of a second decided whether it answered now or waited another round. That check has been moved clear of the boundary, so the pause is the same every time.
After the previous update, conversation mode reliably noticed when you had finished speaking, but it consistently waited about three seconds before replying. That pause was set by how often the speech recogniser re-checked what it had heard, not by the setting that was supposed to control it — so the wait is now roughly 1.8 seconds instead of three. On desktop, a related problem meant the app was not re-checking more often when it needed to; that has been corrected too, so desktop now has the same safety net as mobile when a room is noisy.
In a room with background noise — traffic, a fan, a television, someone else talking — conversation mode could keep listening after you had finished, so you had to tap Finish to send your turn. The app was hearing the room and could not tell it apart from your voice. It now uses a second, independent check: if the speech recogniser stops producing new words while the microphone still sounds busy, your turn ends on its own. It also learns what your room sounds like when nobody is speaking, and adjusts its sensitivity as you use it, so it keeps improving in the places you actually use it. If you are still mid-sentence nothing changes — new words keep the turn open.
When you run several test scripts together, the summary line for the speech and voice suite always read zero passes — even though the suite itself reported every check passing. The suite recorded its totals under different labels than the runner looked for, so the runner fell back to zero. That mattered for more than tidiness: a suite that genuinely failed produced the same zero, so a healthy run and a broken one looked identical in the summary. The labels now match, one other suite was corrected before it could hit the same problem, and a check title that still said “8 entries” while reporting 10 has been reworded.
Three of the optional test scripts you can load from Help › Run Test Script were checking against details that had since changed on purpose — the version number no longer starts with “Beta”, the AI model catalog has been rebalanced several times, the speech voice now starts at a slower default speed, and the app has gained new speech and voice features. Nine checks were reporting failures on a healthy app. Each has been rewritten to check the underlying rule rather than a fixed list or count, so ordinary future changes will not make them wrong again. A duplicated internal setting was also removed while confirming this — it had no effect on the app either way.
When a web page was long enough to be summarised in several pieces, the step that combined those pieces back together ran without the output limits and repetition checks every other combining step uses — so a stuck AI model could fill the result with repeated or endless text, replacing the good summary. That step now uses the same protected path as the rest of the app, and if anything was cut short or had repetition removed the summary says so before you copy it. Found during an internal review of recent changes rather than reported.
Starting a new chat — or opening AI Chat over a notebook that was still loading — could leave the message box unable to accept typing, even though clicking it appeared to work. The box was never given keyboard focus, or lost it a moment later to the page rendering behind. Closing and reopening did not help. Chat now claims focus the same persistent way it already did after creating a note from a chat reply, re-checking a few times as the page settles. Also corrected a scroll bar sizing rule from the previous release that asked very short scrolling areas, including the chat message box itself, for a scroll thumb taller than the area it sits in.
Scroll bars were hard to see — badly so on Dark and Nordic, and on several light themes they were fainter still. Most panes and dialogs had never been styled at all, so they fell back to whatever the system drew; the ones that were styled used a colour meant for hairline borders. Every scroll bar in the app now draws from your theme’s own text colour at a strength measured to stay clearly visible against that theme’s background, so it stands out without looking out of place — and any theme added later is covered automatically. Sizes are consistent too: one width for panes and dialogs, a slimmer one for the toolbar strips, replacing five different sizes.
Three dictation fixes. Choosing a different speech-recognition model had no effect until the app was restarted — dictation kept using whichever model was loaded first, with no indication. Separately, each time dictation paused at the end of a sentence to convert what you had said, the microphone was switched off and back on, and about a second of audio was lost in between; if that pause was triggered by mistake while you were still talking, those words were gone. The microphone is now started again before the old one stops, so nothing is missed. Finally, when the app was busy — syncing, for example — the speech detector could fall behind, mistake the delay for you having stopped talking, and cut a sentence short; it now recognises when it is behind instead of guessing.
The catalog’s global fallback for vision image size went missing three catalog updates ago. Every model currently in the catalog specifies its own value, so nothing you use was affected — but a custom or side-loaded vision model would have quietly used a built-in figure instead of the catalog’s. The value is restored, and the built-in Self Test that caught it passes again.
The Lightweight-tier vision model LFM2 VL 3B is replaced by its successor, LFM2.5 VL 3B, after head-to-head testing: with the recent output safeguards in place, LFM2.5 scored higher on the benchmark suite, and in the notebook OCR test it completed a run that LFM2 could not finish. Same provider (Liquid AI), same license and terms, same download size class. Your default model and the free tier are unchanged. If you had LFM2 VL 3B selected, the app falls back along the standard selection chain; the downloaded LFM2 file can be removed from the AI Models window to reclaim disk space.
The output limits and repetition cleanup added in recent releases only took effect when a document was long enough to be processed in multiple pieces. Smaller documents — including the summary step that combines a page’s notes — took a shortcut that skipped every safeguard, so an AI model could still fill a summary with repeated text without limit. The shortcut now runs through the same protected path as everything else. Found in testing one release after the safeguards shipped; the image-capture protections from that release worked as designed in the same test.
Two fixes to the safeguards added in the last release. First, when an AI model got stuck repeating itself, the cleanup only ever removed the last few hundred characters of the repetition, however long it ran — it now removes the whole repeated section, keeping the real content before it. Second, capturing individual images (as opposed to PDFs or drawings) ran without any of these protections, so a stuck model could fill a note with repeated text and the app would still report success; image captures now use the same guarded engine as PDF pages, and the completion message tells you if any capture was cut short or repaired.
When an AI model got stuck repeating itself while writing a summary, the repeated text was written into the note with no warning — one export ran to 341 pages where only a few were real content. Summaries are now cut at the point the repetition starts, and each model is held to the output length the app already budgeted for it instead of being allowed to generate without limit. Where a summary ends up incomplete, the note now says how many sections were cut short or trimmed, and shorter one-off actions such as Polish or Extract Items report it in the completion message rather than adding text to your note. The Themes and Briefing analyses also now carry an explicit length limit.
When a transcribed page is cut short because the AI model reached its output limit, or repetitive output had to be trimmed away, the page now ends with a clear marker saying so, and the completion message reports how many pages were affected instead of announcing unqualified success. Behind the scenes, the transcription engine now also accounts for each AI model’s actual image-processing geometry rather than assuming one size for all models, which makes cost estimates accurate for every model in the catalog. No change to how pages are transcribed.
Two built-in Self Test checks were writing a problem into the log while still being counted as passes, so neither showed up in the pass/fail total. One was comparing the AI prompt library against an out-of-date count (the library has grown since the number was last set). The other was checking whether the Run Test Script button is available, but it was really measuring whether the Help window happened to be open at the time, so it reported a problem on every run even though the button was correctly available. Both now check the right thing and properly fail if something is genuinely wrong. No app behaviour changes.
The built-in Self Test reported three failures for the GPT-OSS 20B model that was added to the catalog in the previous update. In every case the app and the catalog were correct and the test itself was out of date: it expected an older reasoning-format setting, it treated a deliberately unset sampling value as malformed, and it still listed the model as one that had been retired long ago. The checks now match the shipped, field-tested configuration, and a related check that had quietly stopped running has been restored. No app behaviour changes.
The AI Models window now has a “Check for catalog updates” button. It fetches the latest model catalog from iwvdigitalsolutions.com, verifies its digital signature, shows you exactly what changed (models added or removed), and applies it only if you approve. Models you already downloaded always stay on your device. This update also refreshes the built-in catalog: two new models join — LFM2 VL 3B (a fast vision model under the LFM Open License; free for personal use and for organizations under $10M annual revenue) and GPT-OSS 20B (a high-quality reasoning model for desktop) — while Llama 3.2 3B and Phi-4 Mini are retired from the catalog. Every model download remains integrity-checked against its published SHA-256. This update also cleans up the AI model list: models that carry a license note (the LFM family) now show it in its own box beneath the model entry, so the model name, badges, and memory figures are no longer squeezed.
The August catalog update replaced the previous small default model, but the build configuration still named the old one, which the built-in Self Test correctly flagged. The configured default and all fallback references now point at the catalog’s current recommended model (LFM2.5-VL 1.6B), so the Self Test passes and the app never recommends or falls back to a model that is no longer in the catalog.
Headings that end with a colon — a common style, like numbered principle or section headers — were being excluded from the Table of Contents, leaving some pages listed with no contents at all. They now appear, shown without the trailing colon. Note: AI page descriptions are written when you run the notebook Table of Contents action, so pages added afterwards (for example by transcribing a PDF) show their descriptions after the next run.
Documents with real section headers (short headings over paragraphs of body text, often several sections per page) list those headers in the Table of Contents again, clickable as before. The safeguard from the previous fix remains: content dressed up as headings — like AI transcriptions that tag whole paragraphs or every poem line as a heading — stays out, because headings are only listed when they are a small structural fraction of the note rather than the note itself. A note written entirely as headings lists by its title alone.
The Table of Contents now strictly lists note titles. Headings inside a note’s text (which AI transcription sometimes applies to whole paragraphs or poem lines) no longer appear as entries. Untitled notes get a short AI-written title when you run the notebook Table of Contents action; image notes can be listed by their caption, and image galleries list their name and their images’ captions. Applies to both the combined Table of Contents note and legacy TOC notes.
When a note’s content contained a whole paragraph formatted as a heading (AI transcription output sometimes does this), the Table of Contents listed the entire paragraph as an entry. Headings longer than a sensible label are now treated as content and left out; the note’s own entry remains the way to navigate to it. Applies to both the combined Table of Contents note and legacy TOC notes.
Divider notes and other purely structural items (including a TOC listing itself) no longer appear as entries in the Table of Contents note. Dividers have no title, so each one showed up as a meaningless Divider bullet. Applies to both the combined Table of Contents note and legacy TOC notes in older notebooks.
The Table of Contents and Index note types are now one combined type: a live, clickable contents note with numbered page entries and optional note sub-entries, available from the note picker and the list menu. The notebook-level AI Table of Contents action now updates this note instead of writing a separate text note: it adds a short AI-written description under each page entry (stored with the note, surviving saves and page renames) and creates the contents note on a new front page if none exists. The old approach generated the whole TOC as text, which could not be clicked, went stale, and sometimes came out garbled. Existing Table of Contents and Index notes in old notebooks still work. PDF printing skips the in-app contents note (the PDF builds its own TOC with real page numbers) and gains a new option to include the AI page descriptions under each printed TOC entry.
Two fixes from the same field session. First: asking the AI a question about a PDF note (the custom prompt action) now shows the model the actual pages whenever the current model supports vision, instead of a text-only extraction that scrambled tables and columns — so requests like building a table from an invoice work. Text extraction remains the fallback when the model cannot see. Second: the ’Pages to process’ choice in the Transcribe and Describe dialogs no longer remembers its last setting. It previously restored a stale single-page selection and silently adjusted the page number to fit the new document, which could quietly limit a run to one page. Every dialog now opens set to All pages.
The About > Open Source Licenses screen was missing the current default model (LFM2.5 VL 1.6B) and the largest model (Gemma 4 31B), and still listed two models that left the catalog. The list is now generated to match the shipping catalog exactly, states the LFM model’s licence condition (free for personal use and for organizations under USD 10M annual revenue), and notes which retired models remain covered by their original licences. The in-app Help Guide and the demo notebook now also describe the current default model, the five capability tiers, and the free tier correctly — both free models support image and handwriting recognition. The third-party notices file now travels inside every installed copy of the app.
Using Import in the Restore Backup window with a notebook file (.nxl) or an automatic backup (.nxl.bak) failed with the message ’No notebook found in XML’, even though the file was perfectly valid — including the safety copies the app makes when you close it with unsaved work. The importer was using an outdated reader that could not understand the current notebook format. It now uses the same reader that File > Open uses, so your pages and notes all come through intact. After a successful import, the app also offers to open the notebook immediately instead of only filing it in the backup list.
A new recommended AI model, a vision-capable free tier, and a group of fixes for drawing, saving, and cloud sync. The AI model catalog now leads with LFM2.5 VL 1.6B from Liquid AI — small enough for a phone, yet able to read images and handwriting — and it is available on the free tier alongside Ministral 3 3B. Drawing now stays smooth while a cloud sync runs, cloud sync no longer gives up because you drew something earlier, and notebooks created by dropping files in always save correctly.
If you opened a drawing while a Google Drive or Dropbox sync was running, the sync could quietly mark itself as cancelled and then stop much later — sometimes minutes after you had already closed the drawing — with a misleading message blaming a timeout. The sync now notices when the drawing that paused it has ended and simply carries on. If you are still drawing when it checks, it waits politely and tries again later instead of reporting an error. Real timeouts are still reported exactly as before.
Opening a drawing or PDF annotation while Google Drive or Dropbox sync was running could make the pen skip: strokes were captured at a fraction of the normal rate, and short strokes could be recorded as straight lines instead of curves. NotesXML was already designed to pause sync the moment you start drawing, but the file-comparison step did not check for that pause once it had begun, so it kept working for up to half a minute. It now stops as soon as you pick up the pen.
When the AI is reading an image or handwriting on a phone or tablet, the on-screen notice no longer promises a specific number of seconds. How long it takes depends on the model you chose and the device you are using, and the old estimate was often well short of the real time. The notice now simply says it can take a while and reminds you to keep the app open, and it stays on screen long enough to cover a multi-page job.
In rare cases, a notebook created by dropping files into the app could not be saved — including by the automatic exit rescue — and its contents could be lost when the app closed. These notebooks are now repaired automatically the moment they are saved, and the exit rescue no longer wastes time retrying a save that cannot succeed.
LFM2.5 VL 1.6B from Liquid AI is now the recommended model on both desktop and Android, and it is included in the free tier. It is small enough to run comfortably on a phone while still being able to read images and handwriting, which means image and handwriting recognition is now available without a Professional licence for the first time. Both free-tier models — LFM2.5 VL 1.6B and Ministral 3 3B — are vision-capable. Please note that LFM2.5 VL is offered under the LFM Open License v1.0, which permits commercial use only by organisations under $10M in annual revenue; see the third-party licences page. Llama 3.2 1B and Gemma 4 E2B have been retired from the catalog.
When a device did not have enough free memory to load a vision model, NotesXML could retry repeatedly and leave you waiting without explanation. It now stops promptly and tells you once, clearly, what happened.
After dropping files onto the Pages pane (or a page row), the dashed “Drop to add” overlay now disappears as soon as the drop completes. Previously it could stay on screen after a successful import until the app was restarted.
Diagnostic logs no longer record the internal values used to protect your PIN. If you have previously shared or synced logs from an earlier version, you may wish to delete them from your cloud storage.
This release brings together a large number of improvements since the previous store version. AI analysis of long documents is dramatically faster and now answers multi-part questions reliably, and it keeps the layout of your documents intact. Dictations, transcriptions, drawings and audio recordings are now always saved, even if no notebook was open when you started. Google Drive stays connected instead of disconnecting after about an hour. Drawing and PDF mark-up stay smooth while cloud syncing or auto-saving is running in the background. Settings windows are easier to read, printing gets the cover lines right, and security improvements have been made to diagnostic logging. Very large PDF files (30 MB and up) are now stored in the Media folder beside your notebooks, so notebooks containing them save, reopen and sync normally, and Transcribe finds them. Previously these files were only available until you closed NotesXML. Messages about very large PDFs also now explain clearly what is happening and what to do next. Dragging files onto the editor now adds them to the page you are working on, instead of creating a new page for each large PDF; dropping onto the Pages list creates a page per file, dropping onto empty space in the Notebooks list creates a notebook per file, and dropping onto a specific notebook adds a page per file inside it. On Android, attachments added from the + menu now go on the page you are working on, matching the desktop version; previously each attachment created a new page. Transcribing a whole notebook can now put each transcribed page into its own note: every document gets a new page holding one note per page, which is far easier to read and search than a single long note.
NotesXML keeps a diagnostic log to help identify problems, and that log can be uploaded to your cloud storage or shared when you report an issue. Security improvements have been made to the way information is recorded: sensitive values such as sign-in tokens, passwords and access keys are now automatically hidden from all log entries. Everyday diagnostic detail is unchanged, so logs remain just as useful for troubleshooting.
On Android, Google Drive sync stopped working roughly every hour and asked you to reconnect. NotesXML was renewing your Google sign-in without identifying itself correctly, so Google refused the renewal. NotesXML then treated your still-valid sign-in as expired and discarded it, which is why reconnecting only helped until the next hour. Renewal now identifies NotesXML properly, and a sign-in that is still valid is kept rather than thrown away. Existing users may need to reconnect Google Drive once after updating; it should then stay connected.
NotesXML pauses cloud syncing and auto-saving while you draw or mark up a PDF. A check that clears leftover pauses from a previous session was running about nine seconds after the app opened rather than immediately, so if you started drawing straight away it cancelled the protection you were currently relying on, for the rest of that session. The check now runs the moment NotesXML starts, and only ever clears pauses that were left over from before it started. A related message that wrongly described a normal shutdown as an unclean one has also been corrected.
NotesXML already paused cloud syncing while you were drawing or marking up a PDF, but auto-save was still able to run at the same moment and compete for the same processing, which could cause pen strokes to be recorded unevenly. Auto-save now waits while you are drawing or annotating. It does not wait indefinitely: if you keep working for more than about a minute and a half it saves anyway, because protecting your work matters more than a brief stutter. Saving when you exit is unaffected and still happens immediately.
In taller settings windows — Auto-Save Settings was the reported case — the sections were being squeezed shorter than the text they contained instead of the window scrolling, so one panel was drawn on top of the one above it and both became unreadable. Sections now keep their full height and the window scrolls as intended. This applies to every settings window in NotesXML, not just Auto-Save.
The drop-down lists in Auto-Save Settings took their colours from the application theme rather than from the window they sit in, so on a dark theme they appeared as dark text on a dark background and were difficult or impossible to read. They now use fixed high-contrast colours. Separately, if the Auto-Save Settings window had been opened once before in the same session, a hidden copy of it was left behind, and your choices were read from that hidden copy instead of the one on screen — so a change to the save interval could appear to be accepted while the old setting stayed in effect. The old copy is now removed when the window is opened.
NotesXML pauses cloud syncing while you are drawing or marking up a PDF. If the app was closed unexpectedly during one of those — swiped away from the task switcher, shut down by the system to free memory, or crashed — that pause was never lifted, and syncing stayed switched off the next time you opened NotesXML, without telling you. The pause is now lifted automatically at startup, and also when you exit normally. If a pause is found left over from a previous session, it is recorded in the log so the earlier shutdown can be investigated.
NotesXML pauses cloud syncing while you are drawing or marking up a PDF, so that pen strokes are recorded smoothly. That pause was being reported as though the sync had failed, with a red error message containing technical wording. Nothing had gone wrong, and there was nothing for you to do about it. Pausing is now silent, and the sync indicator simply shows that syncing is paused while you work. Syncing resumes on its own when you finish. Messages about genuine sync problems are unchanged.
If you started a dictation, transcription, drawing or audio recording without a notebook open, NotesXML finished the capture and then had nowhere to put it. For dictation the text was placed on the clipboard and the message said so, but no note was created, and because the clipboard only holds one item, copying anything else afterwards erased the text with no warning. NotesXML now creates a new notebook and page automatically so the note is saved properly, exactly as it already did when you opened a drawing with no notebook. This covers dictation, live transcription, drawings, audio notes and transcripts made from audio or video. If a notebook still cannot be created, the text is copied to the clipboard as before and the message now says clearly that it was not saved.
When Google Drive or Dropbox was syncing in the background, NotesXML could become busy for several seconds at a time comparing files. While that was happening, pen and finger strokes in the drawing screen were only partly recorded, so lines came out broken or missing sections. Marking up a PDF had the same problem and had no protection at all. Drawing and PDF mark-up now pause cloud syncing while you are working, and a sync that is already running stops within a few seconds instead of continuing. Syncing resumes on its own as soon as you finish. Separately, a sync that had quietly stopped making progress could keep extending its own time limit indefinitely; it is now ended once no further progress is measured.
When you asked the AI a question about a long note or document, NotesXML splits the text into pieces small enough for the model to read. Those pieces were being rebuilt with every line break removed, so headings, bulleted lists, table rows, quotations and page markers all arrived as one continuous run of text. The AI could not see where a list or a table began or ended, and answers built from long documents often came back with tables that started correctly and then broke down into a run-on line. Short enough content was never affected, because it was passed through unchanged — the problem only appeared once a document was long enough to be split. The pieces now keep the original line breaks and spacing. Which words go into which piece has not changed, so nothing else about how AI actions run is different.
Text typed onto a drawing note was saved very slightly lower than it appeared in the drawing editor, so it could end up sitting just below a ruled line instead of on it. The drawing editor positions text by the top of the letters, while the saved page positions it by the writing line the letters sit on, and the conversion between the two used a value slightly too large. The gap grew with larger text. NotesXML now measures the correct value for the font in use. As with the previous release, this applies to drawings saved from this version onward.
On a drawing note using the Cornell page style, anything you wrote or drew lined up correctly while the drawing editor was open, but appeared to have moved once you saved and returned to the main view. Your work was never moved — the printed guide lines and labels were. The drawing editor and the saved page were drawing the Cornell layout to two different sets of measurements, so the header band, the cue column and the summary area all changed size and position between the two views. The saved page now uses exactly the same proportions as the editor, including the Topic field and the header underlines that were missing from it. This applies to drawings saved from this version onward; a drawing saved earlier keeps its stored layout until you open and save it again. Other page styles — ruled, graph, dot grid and the rest — were checked and already matched.
The built-in self-test checks that the list of Professional features matches what it expects, as a safeguard against accidental changes to what is included in each tier. Two notebook AI actions — Extract Action Items and Action Plan — were added to the Professional tier in an earlier release, and the self-test’s expected list was not updated to match, so it reported a failure on a correct build. The expected lists are now up to date. A separate internal list of feature descriptions also still mentioned an image-description feature that was consolidated into the vision tools some time ago; that entry has been removed so upgrade messages cannot refer to it. Nothing about which features are included in Free or Professional has changed.
When you ask about a specific chapter, NotesXML narrows its reading to that chapter instead of the whole document. It could only do this when a PDF carried proper bookmarks, or when its chapter titles were printed in a larger typeface. Many documents — particularly plain-text books and older scanned material — have neither: the chapter headings are set in exactly the same type as the body text, so there was nothing to detect and the entire document had to be read. NotesXML now also recognises chapter headings by how they are written, such as a roman numeral or the word Chapter followed by a title. On one 96-page book this reduced a question about Chapter 1 from reading all 96 pages to reading 3. Care has been taken not to mistake numbered footnotes or numbered paragraphs for chapters, and the chapter panel still shows you what will be read before anything runs.
Three fixes to how NotesXML assembles a long answer. First, the final stage that combines partial results was still working in small pieces even after the earlier stages were widened, which is why answers could break off mid-table, state the same fact twice with different details, or trail off into material that did not address the question. That stage now works in the same larger pieces as the rest. Second, when the AI decides a section of a document is not relevant to your question it is supposed to say so and be skipped, but it was only recognised when it said exactly the right word — anything else was treated as part of the answer, which is how a note about excluded material once ended up inside a response. Any clear form of that reply is now recognised. Third, if the AI runs into its output limit partway through a section, the answer now tells you so instead of quietly appearing complete.
The previous release changed how NotesXML sizes the pieces it reads when answering questions about long documents, and stated that no other AI feature would be affected. One safety limit was applied more broadly than intended, which made the Extract Items feature read long notes in about 30 percent more pieces than before — slower, with no benefit. That limit now applies only where it belongs. Extract Items, image analysis and transcription cleanup read long content in the same number of passes as they did before the previous release.
When NotesXML answers a question about long documents, it reads them in pieces and then combines the results. It was cutting those pieces far smaller than your computer could actually handle — often using only about a seventh of the available working memory — which meant many more pieces than necessary. That caused three visible problems: an answer could break off mid-sentence where one piece ended, the same point could appear twice with contradictory details because two pieces were summarised separately, and answers grew padded with material that did not address the question. NotesXML now sizes those pieces to what the model can genuinely hold, reads a document in a single pass when it fits, and avoids cutting in the middle of a bulleted or numbered list. Answers should be shorter, faster and more consistent. Other AI features that read long content in pieces — transcription cleanup, item extraction and image analysis — are deliberately unchanged, because their accuracy depends on their existing reading rhythm.
When you ask about a specific chapter, NotesXML shows you which parts of each document it is about to read and waits for your approval. If it could not narrow things down, the panel simply said “all” — which looked the same whether the document had no detectable chapters, whether your question genuinely covered most of it, or whether the chapter you asked for could not be found. That last case is the one that mattered most: you asked for Chapter 1 and were shown “all”, with no hint that the chapter had not been located. Each situation now has its own plain explanation, including “Chapter 1 could not be located in this document”, so what you approve always tells you why. Nothing about which pages get read has changed — only what the panel tells you.
Two problems could send a document question to the wrong pages. First, referring to a chapter by number — “compare Chapter 1 of the two books” — was effectively ignored: NotesXML matched on the words in your question, and a bare chapter number carried none, so it could land on a different chapter that happened to share a word with the book’s title. Chapter numbers are now understood, in digits, words or roman numerals, and if a number cannot be matched to a chapter NotesXML reads the whole document and says so rather than guessing. Second, some PDFs carry invisible leftover bookmarks from the word processor that created them. NotesXML treated those as the document’s chapter list, which meant it could find no chapters at all and quietly read everything — this affected large manuals in particular. Those leftovers are now recognised and ignored, and NotesXML falls back to detecting chapter headings from the page layout, which finds the real chapters. The chapter panel also now distinguishes between reading a whole document because you asked about all of it, because no structure could be found, and because your question genuinely spans most of it — so what you approve always describes what will run.
When you ask a notebook or page question that names a chapter of an attached PDF, NotesXML shows a panel describing which parts of your documents it is about to read, and waits for your approval. In the previous release that panel could under-describe the work: it might lead with the chapter you named while the actual run quietly included most of the rest of the document as well, because sections with similar-sounding titles were being swept in alongside the one you asked for. The result was a much longer wait and an answer padded with material from chapters you never mentioned. Three things are fixed. First, when you name a chapter, NotesXML now reads that chapter and nothing else — naming the scope means getting the scope. Second, the panel now shows the arithmetic — for example “keeping 14 of 240 pages” — so you can see exactly how much reading you are approving before anything runs. Third, when a question genuinely touches most of a document, NotesXML now says so plainly (“Question terms match most of this document — analyzing full text”) instead of presenting a broad run as a narrow one. Nothing changes for notebooks without PDFs or for the other AI actions.
The previous release added the ability to narrow a notebook or page question down to the relevant chapters of an attached PDF, instead of reading every page of every document. On the normal way of adding a PDF — dragging or dropping the file onto a page, which is how most PDFs get into a notebook — that feature never actually ran. NotesXML looked for PDFs in only one of the two places a PDF can live, found none, and quietly went back to reading everything in full. Because it fell back silently, there was no sign anything was wrong: the panel showing which chapters were about to be read never appeared, and neither did the note explaining that a document had no detectable structure. A long question over a large manual could therefore still take many minutes and come back with a very long answer that wandered well outside the chapters you asked about. NotesXML now finds PDFs whichever way they were added, so the chapter panel appears as intended and you can confirm the scope or choose to read the whole document before anything runs. Each document is also read once and only once: previously, widening the search alone would have meant a dragged-in PDF being sent to the AI twice over. Nothing changes for a notebook or page without PDFs, for the Table of Contents, Key Themes, Briefing and Conflict Analysis actions, or for PDFs that were already being found correctly.
Asking a question about a notebook that contains one or more PDFs used to make NotesXML read every page of every attached document in full before answering — fine for a short document, but on a long manual or report the AI could spend many minutes working through hundreds of irrelevant pages just to find the few that mattered, and the answer could drift into unrelated sections along the way. When NotesXML can find a table of contents or clear chapter headings in an attached PDF, it now works out which chapters or sections your question is actually about — for example, naming a chapter by title, or asking a specific enough question to match a heading — and shows you the chapters it plans to read before it starts, so you can either run it as scoped or tell it to read the whole document instead. Short documents are always read in full, since narrowing them down would not save anything. If a PDF has no table of contents and no headings NotesXML can detect, you will see a note that it is reading the full document rather than a guess at its structure. This only applies to a question you type yourself on a notebook or a page — the Table of Contents, Key Themes, Briefing and Conflict Analysis actions read everything as before, since they summarize the whole notebook rather than answer one question.
When a document was too big to fit the AI’s context window, NotesXML was supposed to ask you how to proceed — summarise it as a note, attach a shortened version, or cancel. On Android that question appeared underneath the chat window instead of on top of it, so none of its three buttons could be tapped and only the device Back button would close it. Worse, if you did not notice the buried question at all, the document was quietly discarded: no attachment appeared in the chat, and the AI then told you it could not see any document. The dialogue now opens above the chat, all three choices work, and it follows your chosen theme instead of always appearing as a white panel. If the question cannot be shown for any reason, NotesXML now tells you plainly that the document was too large rather than dropping it without a word.
One of NotesXML’s internal code files still carried a few lines of leftover web-page markup from an earlier stage of the project. It never affected how the app looked or behaved, but it put a small piece of junk into every assembled build and prevented one of our automated code checks from validating that file. The leftover lines are gone: the check now covers every core file, and builds no longer carry the junk. Nothing changes in how you use NotesXML.
On a Professional Print (and Quick Print with a cover page enabled), the cover always began with the notebook name in large bold type whether you asked for it or not — so the three configurable lines in the print dialog actually came out as the second, third and fourth lines on paper. That automatic line is gone: Line 1 now really is the first printed line, and it takes over the large bold styling the automatic name line used to have. If your saved print settings were made before this change, they are updated automatically so the cover looks the same as before — the notebook name simply becomes your Line 1, and your old Line 1 and Line 2 move down one slot. There is also a new choice in each line’s dropdown, and in the header and footer dropdowns: Filename, which prints the notebook’s file name including its extension. A notebook that has never been saved has no file name yet, so that choice prints nothing until the first save.
Reading a note aloud used to pause only at a full stop, question mark or exclamation mark. Anything without one ran straight on: a heading was spoken as part of the first sentence that followed it, and a bulleted list, a checklist or a table was read as one continuous breath with no gap anywhere. NotesXML now recognises where a paragraph, heading, list item or table row actually ends, and pauses there for twice as long as it pauses between sentences — so structured notes sound the way they look. There is no new setting to learn: the existing Pause between sentences slider in Voice and Speech governs both, and a paragraph simply gets double whatever you have chosen. Setting it to zero still turns all pausing off. A line break inside an address or a verse deliberately does not get the longer pause, because a soft line break is not a new paragraph. Two things that were previously silent are now spoken: numbered lists are read with their numbers, which had only ever existed as on-screen formatting, and a checklist item is announced as done or not done, so listening to a checklist finally tells you what is still outstanding. Reading a whole page or notebook also leaves a paragraph-length gap between one note and the next. The Preview button in Voice and Speech and Conversational Mode are unchanged — they send speech to the engine in one piece and cannot vary the gap within it.
Whenever NotesXML asked you to confirm something — before a long transcription, before overwriting a file, and in several other places — the prompt you saw was the one built into the underlying browser component, not the app. It appeared as a plain white box headed with a web address, ignored your chosen theme, and always showed generic OK and Cancel buttons. It also discarded the wording NotesXML had written for it: the prompt before a long summary was meant to offer "Run it" and "Skip the summary", and you were shown "OK" and "CANCEL" instead. These prompts now use a proper NotesXML dialog that follows your theme and shows the intended title and button labels. The buttons do exactly what they did before, so nothing you were used to has changed meaning. On Android the Back button now closes the prompt the same way as choosing the negative option, matching how it already behaved. Tapping outside a prompt no longer does anything — a question this important should be answered deliberately.
When you transcribe a PDF or handwriting attached to a note, the options screen now shows how many AI passes the job needs — and on Android, roughly how many minutes it will take on your device. The figure updates as you change the page selection, so narrowing to a range shows you the saving immediately, and ticking Executive Summary adds its own cost to the same line. The notebook-level screen has shown this since the previous release; the note-level one did not, which was the wrong way round, since that is where the longest documents usually start. On Android, if the job is estimated to run longer than about fifteen minutes, NotesXML now asks you to confirm before beginning. Choosing to go back leaves the options screen exactly as you had it so you can narrow the page range and try again — nothing is reset. Shorter jobs, and anything on Windows or Linux, start straight away as before.
When NotesXML processes a long document with the local AI, it splits the text into overlapping pieces so nothing is lost at the joins. On phones and other devices where the AI is running with a reduced memory window, those pieces get smaller — but the overlap did not shrink with them, so on a recent phone run nearly half of every piece was text that had already been processed, and the job took roughly twice as many AI passes as it needed. The overlap now scales with the piece size, capped at a quarter of it. On a measured test at reduced memory this took the same document from 51 AI passes down to 24. Full-size settings on desktop are unchanged. Also fixed: an internal safety gap where an unusual combination of settings could have made the splitting step stall indefinitely. Results are unaffected — this is the same work done with less repetition.
Three fixes to professionally numbered PDF exports, all found in one 70-page export. First, the Table of Contents could point to the wrong pages. If your Executive Summary included a page-by-page overview, that overview repeats each page's title, and the contents page was matching those repeats instead of the pages themselves — in the reported export, four entries pointed 24 to 51 pages too early. NotesXML now locates each section by the marker it places at the true start of the section, and it checks that the page numbers only ever increase, since a section cannot begin before the one above it. Second, every page was printing its header and footer twice, slightly offset, which made the date and the page number look doubled or blurred. Only one set is drawn now — the one that accounts for the cover and contents pages and follows your left/centre/right settings. Third, every section began with a blank page. Long notes were being marked as “do not split across pages”, and a note taller than a page cannot honour that, so it was pushed to the next page and split anyway, leaving the page before it empty. Notes now flow normally, while a note's title is still kept with the start of its text so it is never stranded alone at the foot of a page. Expect exports to be a little shorter, and page numbers throughout to shift accordingly.
When you tick Executive Summary (or the other after-processing options) on a transcription or OCR run, NotesXML then works through your notebook with the local AI. On a large notebook that can take a very long time, and until now there was no way to stop it: the progress window showed a single unchanging line, tapping it did nothing, and the Back button did nothing, so the only way out was to force the app closed. That window now has a Cancel button, and Back cancels too. Cancelling stops the run at the next safe point — usually within one step — and everything already finished is kept. The progress line now also tells you which page and which part of it is being worked on and how long the run has been going, so a slow run reads as slow rather than as frozen. Before you start, the estimate on the options screen now includes the cost of the executive summary; on Android, where each step takes far longer, it also shows this in minutes and asks you to confirm before beginning — declining simply skips the summary and runs everything else. Finally, if the AI fails on one note and NotesXML falls back to a shortened extract, the reason is now recorded in the diagnostic log instead of passing silently.
When a notebook file is modified outside NotesXML (for example by a sync service or a file move) and you then save, NotesXML asks how you want to resolve it. That dialog worked correctly, but the app’s diagnostic log never recorded that it appeared or what you chose — which made support investigations harder than they needed to be. The log now records when the dialog is shown and which choice you made. Nothing changes in how saving or the dialog behaves.
Three related Android fixes. First, on phones where the AI was running with a reduced memory window, reading a scanned page could produce an almost-empty result — the app was accidentally counting the page image against its working space twice, leaving no room to write the answer. That double-counting is fixed, so transcription now works in situations where it previously returned next to nothing, and devices with plenty of memory get about a third more room for long answers. Second, if your device is genuinely too low on free memory to load the AI safely, NotesXML now tells you so — “Not enough free memory to load the AI model right now. Close some apps and try again.” — instead of attempting the load anyway, which could cause Android to close the app without warning. If Android does close NotesXML while the AI is loading, you will now see a brief notice explaining what happened the next time you open the app. Third, an internal fix ensures the AI receives images in the intended way when using the default Ministral models.
The 5.171.4 checklist print fix only worked for checklists in one internal storage format, and notebooks you had already saved use a slightly different one — so your printed checklists still showed the text on the line below the checkbox. The print styling now recognizes both formats, so checklists in every notebook — old and new — print with the checkbox and its text side by side and completed items struck through.
When you printed a notebook or note to PDF, checklist items came out wrong: the checkbox appeared on one line and its text dropped to the line below, and completed items lost their strike-through. On screen everything was fine — the problem was only in the printed document, which uses its own styling that had no rules for checklists at all. Printed checklists now match what you see on screen: checkbox and text side by side, and checked items shown with a filled box and a line through the text, so completion status survives even in black-and-white printing. Also fixed: when a note has the same title as the page it sits on, the printed page no longer repeats that heading twice in a row — you will see the page heading once, followed by the note content.
On some devices with very limited memory, the AI could return an almost-empty answer for a scanned page — sometimes a single word — because it ran out of working space before it could write anything. NotesXML accepted these near-empty answers as successful transcriptions: the automatic retries you configured never ran, the completion message reported every page as transcribed, and a note containing almost nothing was created as if it were the real text. NotesXML now recognizes this ran-out-of-space condition, retries the page the number of times you asked for, and if it still cannot get a real transcription, honestly reports that page as failed instead of saving a stub. Pages that genuinely contain only a few words — a chapter title, a page number — are still transcribed and saved exactly as before.
AI models sometimes wrap an entire page's transcription in code-fence characters — decoration the model adds, not text from your document. A cleaner already existed to remove that wrapper, but on multi-page transcriptions the page label was attached to the text first, which hid the wrapper from the cleaner. Affected pages then displayed as a single block of computer-style text — literal # and ** symbols and vertical bars in a monospaced font — instead of proper headings, bold text, and tables. The wrapper is now removed at the moment the transcription is received, before the page label is attached, so every transcription path shows the page's real formatting. Backticks that genuinely appear on a scanned page are still preserved exactly. Already-saved transcriptions are unchanged; re-run the transcription (tick Re-do in Transcribe & OCR Notebook, or run Transcribe Text again) to regenerate any affected pages.
During OCR, a safety check watches for the AI getting stuck repeating itself, and trims the output when it does. That check mistook ordinary table structure — the aligned spacing between columns and the dashed separator lines — for the AI repeating itself, so pages containing tables were cut short and ended with a notice that repetitive content was omitted; in one real document both tables were lost this way. The check now understands that table structure and repeated cell values such as identical prices are part of the document, not a malfunction, and when the AI genuinely does get stuck, the trim now keeps everything up to where the repetition actually began instead of cutting from the first place a phrase appeared. Re-run transcription with Redo ticked to recover any pages that were previously cut short.
When Transcribe & OCR Notebook reads a document that contains a table, the result note now shows a proper table instead of rows of text separated by pipe characters. This applies when the Output detail choice is Formatted (the default), which already asks the AI to keep headings, lists, and tables; the note now displays that structure rather than the raw markup. Choosing Faithful still produces exact character-for-character text, for documents where literal symbols matter. Notes transcribed before this version keep their old appearance until you re-run transcription with the Redo option ticked.
Until now, only the Exit button in the File menu checked for unsaved notebooks before quitting on desktop. Closing the window with the title-bar X, Alt+F4, or a system shutdown skipped that check entirely — the window simply closed, and any unsaved changes were lost. Closing the window now runs the same safety path as the Exit button: unsaved notebooks are saved first, and if anything cannot be saved you are told and given the chance to stay. If the app ever hangs during this check, pressing the close button a second time forces the window shut immediately, so you are never trapped. Android is unaffected — it already had its own exit handling.
Clean Up Formatting includes an instruction to collect any stamps, classification banners, or reference numbers a document already carries into a section of their own. The instruction was worded as a command rather than a condition, and on documents that had no such markings the AI would sometimes obey it by making them up — in one real case inventing a confidentiality classification and page numbers for an ordinary lecture transcript. The instruction now applies only when such markings genuinely exist in the text, collects them word-for-word, and explicitly forbids inventing any. If you have customized the Clean Up Formatting prompt in the Prompt Editor, your customized version still contains the old wording — reset that prompt to its default to pick up this fix.
When the AI works on a document too long to handle in one pass, it processes it in several pieces and joins the results back together. The joining step had a long-standing fault: it rebuilt each piece as a single run-on line, discarding every line break — so headings, lists, tables and paragraph breaks produced by the AI were flattened, and a short run of words could be lost where two pieces met. This was the reassembly fault behind the withdrawal of whole-notebook Clean Up Formatting in 5.169.3. The joining step now preserves line structure exactly, and Clean Up Formatting on a single note additionally skips the overlap trimming that caused the word loss. This repair is one prerequisite for the whole-notebook option's return, but does not by itself restore it.
The Clean Up Formatting option added to Transcribe & OCR Notebook in 5.169.0 has been removed. In real use on a long document it could scramble the transcription it was meant to tidy: on documents long enough to be processed in several pieces, the reassembly step could run paragraphs together, drop short runs of words at the joins, and leave stray formatting marks mid-document. The transcription itself was not the problem — the raw text captured from your pages was good, and it remains untouched when this option is absent. Clean Up Formatting on a single note is not affected and works as before. The whole-notebook version will return once the reassembly step can guarantee the wording and line structure survive intact. The Output detail choice (Formatted or Faithful) added in 5.169.0 remains, as does the fix that gave PDF pages in notebook runs the same formatting instructions as single-note runs.
NotesXML already refuses to attach files that a computer would run directly, such as programs and shell scripts, so that a notebook cannot become a way of passing one around. Eleven more types have been added to that list: .desktop and .appimage on Linux, and .hta, .msc, .cpl, .scf, .inf, .chm, .wsc, .sct and .jnlp on Windows. Each of these is run by the system when opened, rather than simply viewed. Attachments already saved inside existing notebooks are not affected and can still be downloaded. Please note this check looks at the file type only, so it cannot see inside a zip or other archive; treat archives from people you do not know with the same care you would anywhere else.
Three AI tools added in an earlier release were broken from the moment they shipped. On a drawing or handwriting note, Describe Drawing produced a short title instead of a description. Auto Title (Vision) and, on image galleries, Caption Images were missing from the AI menu altogether, so there was no way to reach them. All three now behave as intended: Describe Drawing describes the drawing and offers the page picker on multi-page notes, Auto Title (Vision) appears in the Drawing AI list and names the note from its first page, and Caption Images appears in the Gallery AI list. The counts shown next to each collapsed section were also short by one and now match what is actually listed.
Transcribe & OCR Notebook now offers the same two output controls the single-note OCR has always had. Clean Up Formatting, a post-processing option, reorganizes each transcription into clean structure — headings promoted, lines broken by a page edge joined back into paragraphs, columnar and form layouts turned into tables — without changing any of the words. It is switched off by default, and when you switch it on the estimate at the top of the window updates to show the extra work involved. Output detail lets you choose Formatted, which gives transcriptions clean structure, or Faithful, which reproduces only the structure visibly present on the page. Running the action again does not redo formatting that has already been cleaned up unless you tick Re-do. This release also corrects a related fault: PDF pages transcribed as part of a whole-notebook run were being read without the formatting instructions used everywhere else, so they came out noticeably flatter than the same PDF transcribed on its own. They now match.
On a fresh installation, the one-time acknowledgment about AI-generated content could appear underneath a progress window — most visibly when running a notebook transcription — leaving the app waiting on a dialog you could not see or tap. Two changes fix this for good. First, the acknowledgment now appears the first time you open the AI Models screen or download any AI model, including voice recognition models — on a clean screen, before anything else is running — so it can never again collide with a progress window. You only ever see it once. Second, the underlying display-order fault that let other windows cover it has been corrected, so even users updating from an older version, who already have models installed, are protected.
Text to Speech in AI Chat and Reading Mode produced an audible click when audio first started and another when it stopped. The cause was a background helper that keeps the audio device awake between sentences: it was feeding the device a signal in the wrong audio format, one where the value it sent means ‘loudest possible negative level’ instead of silence. The level itself was inaudible, but the jump into it and out of it produced the clicks. The helper now uses a format where that value truly is silence, so both clicks are gone. It also now matches the audio system’s native sample rate, removing a device mode-switch that could add its own artifacts.
The default Text to Speech speed was found to be uncomfortably fast for extended listening. New installations and profiles that have never adjusted the Speed slider now start at 0.8x instead of 1.0x, and the Reset button in Voice Settings returns to 0.8x. If you have already chosen a speed yourself, your setting is kept exactly as it is — defaults never override an explicit choice.
The AI menu for a notebook now includes Auto Title & Caption Images, which walks every page and gives each image note a concise title and a one-sentence caption, and writes a caption under every photo in your image galleries. Items that already have a title and caption are skipped, so re-running it after adding new photos only processes the new material — clear a title or caption to have that item redone. The matching page-level action has also been extended: it previously skipped image galleries on the page without saying so; it now captions them too.
Two page-level AI actions are now available at the notebook level. Extract Action Items gathers every task, decision, deadline, and open question from all pages into one organized document, with each item showing which page it came from and duplicates merged. Action Plan turns the whole notebook into a single prioritized plan. Both are Professional-tier, matching the other notebook analyses.
Drawing and handwriting notes gain Auto Title (Vision), which looks at the drawing itself to produce a title — something the text-based Auto Title could never do, since a drawing has no text layer. Image galleries gain Caption Images, which writes a descriptive sentence under each photo; it sits beside the existing OCR & Title Images, which writes short titles in the same place, so you can choose either style.
When you exited the app with unsaved work, NotesXML could show a message saying your notebooks had been saved when in fact some of them had not been written to disk. Two separate faults combined to cause it. First, a single tap on Exit was being handled twice, and the second attempt collided with the first while it was still saving — the collided saves were then counted as successes. Second, the exit routine never checked whether each save had actually completed, so a save that was refused was still reported as done. Both are fixed: one tap now runs one exit, and the exit message counts only saves that genuinely finished. If a notebook cannot be saved, the app now says so plainly, names it, and suggests cancelling the exit so you can save it yourself. A related change stops the exit process from opening a Save As dialog it has no time to complete — that notebook is now reported as unsaved instead, which is the honest answer.
Notebooks created by importing an Evernote (.enex), CSV, Markdown, or plain text file were internally marked as though they already had a file on disk, when they did not. The app could therefore try to open a Save As dialog at an unexpected time — including while closing down. Imported notebooks are now correctly treated as new until you save them for the first time, at which point you choose the location as normal. The notebook name shown in the header is unchanged.
The demo notebook, Help Guide, About dialog, and Open Source Licenses list all described an older version of the AI model catalog. They now match what actually ships: ten models across six tiers, with Ministral 3 3B as the recommended model on Windows and Linux and Llama 3.2 1B on Android. Two models that are no longer offered have been removed from the licenses list, one that was missing has been added, and the llama.cpp engine version shown is now the one actually bundled.
The AI model licensing section of the in-app EULA listed a model that is no longer offered and omitted three Mistral AI models that are — including Ministral 3 3B, which is available on the Free tier. That section now matches the model list published on our website, and the note-type count stated in the feature table has been corrected. No terms affecting your rights, your data, or your license have changed.
An internal safety check that repairs damaged notebook files was mistakenly being applied to every kind of file being opened, not just standard notebook files. Two kinds of perfectly good files were harmed by it before they could be read. First, notebook files in the older multi-notebook format could not be opened at all: the repair step cut off part of the file and the open then failed with a misleading message saying the file was corrupted -- it never was. Second, a document that merely quotes notebook markup inside it -- for example a Markdown file with a code sample -- could open with everything after the quoted sample silently missing. The repair step now first checks what kind of file it is looking at and only touches standard notebook files, exactly as it always has for those; other file types are passed through untouched to be read normally. No files on disk were ever damaged by this issue. This release also adds a large automated test suite that exercises every way a file can be opened -- picker, recent files, drag and drop, double-click, startup restore, and the Android import paths -- with real files of every supported format, which is how this issue was found.
When an AI action such as Executive Summary produced a comparison table, the table could appear as rows of raw pipe characters instead of a formatted table -- in the note and in the PDF export. The cause: the AI model sometimes wraps long table rows across several lines, which stops the table from being recognized. NotesXML now detects and rejoins these wrapped rows before converting the text, so the table renders properly; the table content itself is never altered, and if a passage merely resembles a broken table it is left exactly as written. Rendered tables in PDF exports are also now kept within the page width, with long cell content wrapping instead of running off the page.
NotesXML already cleaned invisible and look-alike characters out of text written by the AI before saving it into your notes — these are characters you cannot see on screen but which travel with the text when you export or share a note. That cleanup was not reaching everywhere it should. Two places were missed: AI replies that arrive word by word as they are generated (which is how the AI Chat panel and, on Android, the punctuation and paragraph cleanup after dictation both work), and one older connection path used when NotesXML talks to a separate AI server. Both are now covered, so the same output is cleaned the same way no matter which path produced it or which platform you are on. If your AI output was already clean, which is the normal case, nothing about it changes.
The Vision section of the AI Prompt Editor gains a new entry: PDF OCR. This is the instruction used when transcribing PDF pages — both when you run Transcribe Text (OCR) on a PDF note and when Transcribe & OCR Notebook processes PDF sources. Previously this prompt was built in and could not be changed; editing it now applies to every PDF page read, in both places, from one setting. Until you edit it, behaviour is exactly as before, and the Reset button restores the standard prompt at any time. (The related Transcribe Text (OCR) prompt continues to govern images and photos.)
If you typed in a note and then immediately clicked over to another notebook, your last edits could be silently lost: the automatic save that normally runs when you leave the editor was still waiting its turn when the notebook switch happened, and by the time it ran, the previous notebook was no longer on screen to save from. The switch itself now saves any pending edits first, before anything changes, so what you typed is always captured — no matter how quickly you switch. Switching without editing writes nothing and will not mark a notebook as changed or trigger a sync upload.
In the AI Prompt Editor, the row of prompt names at the top of the window could hold more prompts than fit across the dialog — the ones past the edge were cut off and could not be selected or edited. The list now wraps onto multiple lines and scrolls vertically when a category has many prompts, so every prompt in every category can be reached, selected, and edited on both desktop and Android. Also includes internal code cleanup with no behaviour change.
The page-level OCR All Images and OCR & Describe All actions used to include multi-page drawing and handwriting notes, but could only read the first few pages of them in one pass — so a long handwritten note came back silently incomplete. Those notes are now left out of the page-level sweep, exactly as PDFs always have been, and the result includes a short notice telling you which notes were skipped and to use Transcribe Handwriting on the note itself, which reads every page properly with progress and retries. Single-page drawings are unaffected and still included. If a page contains only multi-page drawings, the page-level OCR options are not offered, since there is nothing there they could process completely.
We have aligned which AI actions are free and which require Professional so the same capability is treated the same way on every note type. The rule is simple: describing a picture is free; reading the text out of a picture is Professional. What stays free: Describe Drawing, Describe Image, and Auto Title & Caption. What now requires Professional (matching how these same capabilities already worked on PDF and image notes): Summarize Drawing, Convert to Outline, Solve Math, and List Diagram Elements on drawing notes, and the vision-based Summarize and Outline on PDFs and images. Translating handwriting or page images was already Professional and is unchanged. If you were using one of these actions on the free tier, we apologize for the change — it closes an inconsistency where the identical feature was paid on one note type and free on another, which was not sustainable as we head into paid distribution. The regular Summarize on typed text remains free everywhere.
Previously, after saving or reopening a notebook, running Transcribe & OCR Notebook again would re-process every item — even ones already transcribed and unchanged — because the internal record of finished work did not survive the save. That record is now stored durably with each companion note in the notebook file itself, so a re-run skips everything that has not changed and only processes what is new or edited. Notebooks saved by this version remain fully compatible with earlier versions: older versions simply will not skip already-finished work. Editing a companion note’s text no longer affects this either — your edits are yours, and the transcription record is kept separately.
Occasionally the AI wraps a transcription in code-fence characters (three backticks) and then adds a short note of its own after the closing fence — for example a remark about how the page was read. Previously that pattern kept the stray fence characters in your note. The wrapper is now removed in that case too, and the AI’s closing remark is kept as normal text. As before, genuine code blocks and any fence characters that are truly part of your content are left exactly as written.
While transcribing a notebook, the progress dialog could briefly show “page 2 of null” during a retry, and the counter underneath read “Page 0 of 11” even though it was counting sources, not pages. The page total now always shows the real number, and the counter says what it means — “Source 3 of 11” — advancing as each note finishes.
Three improvements from field testing. First, the Table of Contents option now creates the standard clickable Table of Contents note — the same one available from the note menu, where clicking any entry jumps straight to that section — placed on its own Home page at the front of the notebook. It builds instantly with no AI call and stays current as your notebook changes. (Previously this option generated an AI-written text outline, which could be slow and uneven; that AI outline is still available from the notebook AI menu as Table of Contents.) Second, an Executive Summary checkbox has been added: tick it and a full notebook summary page is created after transcription finishes, using the same reliable process the PDF tools use. Third, the dialog itself now looks and works like the PDF page-selection dialog, with the same After-processing options presented the same way. Your choices are remembered for next time.
Two problems from real-world testing of the Transcribe & OCR follow-up actions. First, the completion message could appear while the table of contents was still being generated in the background — in one case seventeen minutes early. Closing or restarting the app on the strength of that message silently discarded the unfinished work. The follow-up actions now genuinely complete before the summary message appears, and the elapsed time is accurate. Second, the generated table of contents used to require an extra click in a results dialog and was then saved onto whichever page you happened to be viewing — easy to miss entirely. When run as a transcription follow-up it now creates its own page at the front of the notebook, named Table of Contents, with no dialog. Running Table of Contents directly from the AI menu behaves exactly as before.
In an installation where one internal component failed to load, Transcribe & OCR Notebook reported every item as failed instantly, with no explanation. It now checks that its transcription engine is actually present before showing the dialog, and if it is not, tells you plainly and suggests a restart. If individual items fail during a run, the completion message now includes the first error so a problem can be diagnosed instead of guessed at.
The new transcription dialog appeared as a bright white panel regardless of your chosen theme — jarring on any of the ten dark themes. It now detects your theme and renders with matching colors, like every other dialog in the app. The input field and checkboxes remain clearly visible on all 16 themes.
The Transcribe & OCR Notebook dialog now has three optional follow-up actions: Title Pages, Notebook Title, and Table of Contents. Tick any of them and they run automatically after the transcription finishes — which is exactly when they work best, because by then the AI can actually read your handwriting through the new transcription notes. They run in the right order (pages are titled first, since the notebook title and the table of contents both use those titles), your choices are remembered for next time, and if everything is already transcribed you can still run just the follow-ups — for example, to build a table of contents over transcriptions you made earlier. Also fixed: the AI sometimes wraps a transcription in code-fence characters (three backticks) that were never on your page; these stray markers no longer appear in transcription results. Text genuinely containing such characters is left exactly as written.
The notebook AI menu has a new action: Transcribe & OCR Notebook. Until now, making a mixed notebook readable meant visiting every page and every note separately — one command for handwriting here, another for a PDF there, another for each image. This action walks the whole notebook for you. It first shows you exactly what it found — how many handwriting notes, PDFs and images, how many pages, and roughly how many AI calls the run will take — so you can decide before anything starts. Each transcription is placed as a new text note directly beside the note it came from, so the content lives where its source lives. You can cancel at any time and keep what finished, and running it again skips anything already done — only new or edited notes are re-read. A bonus honesty fix: notebook analyses like Executive Summary used to quietly ignore handwriting and images they could not read; they now tell you when sources were missed and suggest transcribing first. Once transcriptions exist, those analyses automatically include them. Requires NotesXML Professional and a vision-capable AI model.
Until now, Summarize and Translate on a PDF worked from the document’s embedded text — which a scanned document or a photographed page simply does not have, so the only route was to transcribe first and then summarize the transcription. PDF and image notes now have three new actions in their Vision menu — Summarize (Vision), Outline (Vision), and Translate (Vision) — that read the page image itself, so a scanned archive or a photo of a whiteboard can be summarized in one step. On multi-page PDFs you choose which pages to include, each page is read separately with progress and cancel, and the per-page results are combined into a single summary or outline at the end. Scan Receipt on a multi-page PDF now also reads every page instead of stopping after the first few, giving you one extraction per receipt. Also fixed: the “Cleanup formatting” option in the PDF page-selection dialog was being ignored — it now runs when ticked. All three new prompts can be customised in Settings under AI Prompts, Vision.
Occasionally an AI result — an executive summary, for instance — arrived looking like unformatted computer text, with visible ## marks instead of headings and a table that never became a table. The AI had wrapped its whole answer in a code marker, and NotesXML took it at its word. That wrapper is now recognised and removed, while genuine code — yours, or code the AI wrote for you — is left exactly as it is. Separately and more importantly: when one of those blocks, or any code block, was exported to PDF, lines that were too wide for the page were simply cut off at the right edge, with no warning that words were missing. Code blocks in PDF exports now wrap onto the next line instead, so nothing is lost. Notes that were already saved in the broken wrapped form are not rewritten — re-run the AI action to regenerate them cleanly.
Asking a multi-page handwritten note to summarise itself, build an outline, solve its maths, describe itself, translate itself or list its diagram elements only ever looked at the first few pages — often just four — and quietly stopped there. Transcribe Handwriting was the one action that already read every page. All of them now do. Each page is read separately, so there is no longer a limit on how long a note can be, and you get a progress display, a Cancel button, and anything already finished is kept if you stop early or a page fails. Where it makes sense, the per-page results are then combined into a single result: a summary becomes one summary rather than several. Where it does not make sense, they are kept separate — transcriptions are never rewritten, and a note holding several receipts gives you one result per receipt. You can also choose which pages to process, and how many times to retry a page that fails. Single-page drawings are unchanged and still run in one quick pass.
Opening the AI menu on a drawing or handwriting note used to show just four general actions, with everything specific to drawings tucked away behind a closed “Show Drawing AI” line you had to notice and click. Those actions are now open on the menu where you can see them. Three capabilities that were previously only available on PDFs and images have also been added for drawings: OCR + Image Description, which transcribes a page and describes any sketches in place; Scan Receipt, for handwritten receipts and expense logs; and Analyze Diagram, which explains how the parts of a hand-drawn diagram relate to each other. The existing “Identify Diagram” action has been renamed “List Diagram Elements” so it is clear how it differs from Analyze Diagram — it lists what is in the drawing, while Analyze Diagram explains how it fits together. If you had customised the wording of that prompt, your version is kept. Transcribe Handwriting and Describe Drawing were already tuned for hand-drawn content and have not been duplicated. On very tall menus the Vision group is the part that now folds away instead.
Creating a note, typing into it and pressing Read produced “This note has no text to read”. Read Aloud was looking at the last saved copy of the note rather than what was on screen, so anything typed since the last save was invisible to it. It now reads what you can actually see. Notes on other pages, which are not on screen, are read from the saved copy exactly as before, and a note that really is empty still tells you so. Separately, paragraphs are no longer run together when a note is read aloud.
The new “Pause between sentences” slider did not work: dragging it left the reading unchanged and the value beside it never moved. Two controls in the Voice & Speech panel were accidentally sharing the same internal name, so the sentence-pause slider was quietly updating the Conversational Mode setting instead of its own. Both now work independently. The pause also now applies to the Preview button and to Conversational Mode, which previously ignored it.
Read Aloud ran sentences together with no real gap, which sounded rushed and unnatural on anything longer than a line or two. Voice & Speech settings now has a “Pause between sentences” slider, sitting alongside Speed and Pitch, which lets you set anything from no pause at all up to a full second. It defaults to a third of a second, which most people find natural for reading a page of notes. The setting applies on Windows, Linux and Android, takes effect on the next sentence without restarting, and is included in the Reset button. On desktop, the short delay before the very first word is left alone, since that exists to let speakers and Bluetooth devices wake up before you miss anything.
On desktop, the transcription window closed as soon as the speech had been recognised — but NotesXML was not finished. It still had to tidy the punctuation, which on a two-minute recording took another eleven seconds, and during that time nothing at all was on screen before the note finally appeared. The progress window now stays up, with the turning circle, until every step is genuinely done. Android was not affected by this and already behaved correctly.
Four things about transcribing a recording have been put right. The progress window used to tick every step as finished, including “Complete”, while NotesXML was still tidying up the punctuation — which on a recent recording meant thirty seconds sitting in front of a window that looked finished but was not. Progress now stays on the formatting step, with a turning circle to show work is still going on, and only says Complete when it genuinely is. Second, the message that appeared when the AI decided not to change your punctuation said it could not format the text, right after it had in fact split the transcript into paragraphs; it now says plainly that punctuation was kept as transcribed and paragraph breaks were still applied. Third, the timestamps down the side of a transcript were wrong on Android — a two-minute recording could show times running into the hours. They are now correct, and so is tapping one to jump to that point. Finally, where the speech recognition marks a change of speaker, NotesXML now starts a new paragraph instead of leaving the raw “>>” marks in your text.
If the AI model saved in your settings was no longer offered in the current model catalogue — which could happen after a catalogue change or an update — NotesXML quietly lost track of it. No model appeared selected in the list, and pressing Download reported “No model selected”, which was misleading: you had chosen a model, and the app had simply lost the ability to find it. On Android there was no way to recover apart from noticing the problem and picking another model by hand. NotesXML now checks your saved model against the catalogue when the AI Models screen opens, and if it is missing it selects the recommended model for your device and remembers that choice. The Download button also no longer sits there looking usable when there is nothing behind it — it now tells you to choose a model instead of failing when pressed.
Two things went wrong while setting NotesXML up for the first time. On the AI Models step, pressing Download without first tapping a model reported “No model selected” — but nothing had ever looked selected, because on a brand-new install there is no saved choice to show. The recommended model for your device is now selected for you when you arrive at that step, so Download works straight away; you can still pick a different one. Separately, if you downloaded a second speech-to-text model during setup — Whisper Small alongside the standard one, say — there was no way to switch to it without finishing setup and going to Settings afterwards. The Set Active button now appears in setup as well, so you can choose which voice model to use right where you downloaded it. Delete is still kept out of setup on purpose.
The first time you transcribed a recording, NotesXML could appear to freeze just as it finished. The progress window stayed on screen with every step ticked off, and behind it — invisible and out of reach — sat the one-time notice about AI-generated content, waiting for an answer it could never receive. The only way out was to press Cancel Transcription on a recording that had in fact already been transcribed successfully. NotesXML now asks you to acknowledge that notice before transcription begins, while nothing else is on screen, so the two can no longer collide; if you decline, your recording is still transcribed, just without the automatic punctuation tidy-up. The same protection has been extended to the other places a progress window can be showing when the notice appears. The Cancel button also now disappears once all steps are complete, instead of inviting you to cancel finished work. This only ever affected the first transcription after installing, and no recordings or transcripts were lost.
NotesXML now records voice notes in a more efficient format. On Android the recorder uses Opus where the device supports it, and a more efficient AAC setting where it does not. A four-hour recording that previously took about 230 MB now takes roughly 58 MB, so long recordings take up far less room and sync much faster. Speech-to-text accuracy is unchanged — transcription always works from a separate high-quality stream and is not affected by the recording format. Recordings you already have are unaffected and still play normally.
NotesXML now sets a clear limit on files stored inside a notebook. A single attachment kept inside the notebook can be up to 10 MB; on desktop, larger files you drag in are stored alongside the notebook instead, so they no longer bloat it. NotesXML also watches the notebook as a whole and will warn you when it is carrying a lot of embedded files, and will stop you adding more once it gets very large — big notebooks are slow to open, save and sync. On Android, attaching a file from the file picker previously had no size limit at all, which could make a notebook impossible to open; that is now fixed. Photos, audio and video notes are unchanged. Files already in your notebooks are not affected.
NotesXML already refused to open program and script files that were attached to a note — installers, batch files, shell scripts and similar. Those files can now no longer be attached in the first place, whether you pick them with the attach button, drag them onto a page, paste them from the clipboard, or bring them in as part of a folder or an Evernote import. If you try, NotesXML tells you which file was refused and why, and everything else in the same batch is still added normally. Files already attached to your existing notebooks are not affected in any way and can still be downloaded as before. Archives such as .zip are still allowed.
Tapping Open on a file attached to a note failed with a message about a temporary file that could not be created. NotesXML was building the temporary location incorrectly, so the file was never written anywhere it could be opened from. This affected every kind of attachment, not just images — PDFs were unaffected only because they use a different route that already had the fix. Attachments now open normally, and if the temporary folder genuinely is unavailable you get a clear message saying so.
When you opened a notebook by tapping it in a file manager and then saved it, NotesXML could not tell Android which file to write back to. Depending on what else was open, it either saved a second copy called something like ’My Notes (1).nxl’ or, in the worst case, wrote over a different notebook. If you had just added a password, that also meant your unprotected original was left in place. NotesXML now finds the real file inside your NotesXML folder and updates it directly. For a notebook stored outside that folder it asks you where to save instead — it will never guess.
On Android, opening a notebook by tapping it in a file manager gave NotesXML a temporary working copy without remembering where the real file was. Saving — including saving after adding a password — therefore created a NEW file with a strange ’temp_open_’ name instead of updating your notebook, and if you had just encrypted it, the original file was left unprotected beside the new one. NotesXML now remembers the real file. Where Android permits it, Save updates your original in place; where it does not, the save dialog at least suggests the correct name. The Recent Files list can also reopen these notebooks now. If you encrypted a notebook before this fix, check for an unprotected original and delete it after confirming the encrypted copy opens.
On Android, opening a notebook from a file manager or another app showed an internal working name in the title bar, such as “temp_open_My Notes.nxl” instead of “My Notes.nxl”. The correct name is now displayed. The notebook itself was never affected. Also corrected: a harmless internal message was written to the log every time a notebook was opened, reporting a problem that was not one. No notes were ever at risk from it.
On Android, opening a notebook that arrives through the import bridge — which is how most .nxl files reach the app from other programs and file managers — used a separate, older routine of its own. That routine had no way to ask for a password, so an encrypted notebook simply failed to open. It now uses the same shared open routine as every other part of NotesXML, so it prompts for the password, repairs a notebook damaged by an interrupted save, and recovers properly if a very large file stalls. Imported notebooks also no longer record a filename in the place a full file location belongs.
Three of the ways NotesXML opens files each carried their own private copy of the open logic, which had drifted apart over time. They now all use the same shared routine, so every method benefits from the same format detection, password prompting for encrypted notebooks, and Recent Files behaviour. Two real problems are fixed along the way: files imported on Android no longer record a filename where a full file location belongs, and a notebook saved by a future version of NotesXML will no longer be misread as a very old format.
Every time NotesXML started, it recorded several messages marked CRITICAL about a list button that could not be found. Nothing was actually wrong: that button was replaced by the formatting ribbon some releases ago, and the leftover start-up code was simply looking for something that no longer exists. The leftover code has been removed, so those misleading messages are gone. Creating bullet, numbered, and checkbox lists is unchanged.
The previous release said development and beta builds would write their full internal diagnostics to the log file. They did not: the detail was being sent to an internal logging component that has been switched off since an early Alpha, so nothing reached the file. Those messages now go through the logging path that is actually in use. Released versions remain unchanged.
Development and beta builds now write their full internal diagnostics to the log file. Previously a large share of that detail was only visible inside the app’s debug console, so it was missing from any log sent in with a problem report — which made some issues much harder to track down. Released versions are unchanged and keep the same, smaller logs.
When you launched NotesXML by double-clicking the notebook you had open last, the app loaded that file a second time as part of restoring your previous session. You saw one notebook and nothing was lost, but the file was read and parsed twice, which slowed startup noticeably for large notebooks. NotesXML now recognises that the file is already open and skips the redundant reload.
Log files record file names, folder paths, and account details in plain text, so they are not safe to send to anyone without review. The shortcuts that open or reveal them have been removed from released versions of NotesXML and now appear only in development and beta test builds. Logging itself is unchanged and still runs normally. If you need to report a problem, use Generate Bug Report in Help & Info — it collects what we need with personal details removed automatically.
Pressing the log item in Help & Info did its work, but on desktop the file manager window could open BEHIND NotesXML, so it looked like nothing happened. The window now comes to the front with the newest log highlighted. The item is also renamed from Share Latest Log to Show Latest Log File to better describe what it does — on Android it continues to open the share sheet.
The Open Logs Folder and Share Latest Log items added in the previous release were wired to an older Help menu that the app no longer shows. They now appear in the Help & Info panel under Diagnostics, right beside Generate Bug Report — on desktop you will see both; on Android you will see Share Latest Log.
The Help menu now has two new items right beside Send Feedback: Open Logs Folder (desktop) takes you straight to where NotesXML keeps its log files, and Share Latest Log opens your device’s share options with the current log — on Android this uses the standard share sheet, and on desktop it highlights the newest log file for you in your file manager. When you share, the very latest lines are saved to the file first, so nothing recent is missing. Handy when reporting a problem.
Three refinements from field testing. First, if a file in your cloud storage has a name too long for this device to store, sync now skips it gracefully with a clear one-time notice instead of failing over and over in the background — the file stays safe in your cloud storage. Second, opening the built-in demo (or any notebook that has never been saved to disk) no longer writes needless error lines to the log. Third, exported task lists showed a garbled character where the checkbox symbol should be — the proper empty-checkbox glyph is back, and an internal audit gate now guards against the invisible-character corruption that caused it.
Clicking a Recent Files entry whose file has been moved or deleted now tells you plainly that the file no longer exists and removes the dead entry — previously this also wrote a confusing empty error to the log. If a file fails to open for any other reason (for example a permissions problem), the entry now STAYS in your list instead of being deleted, and the message tells you what went wrong and what to try. Also fixed: the built-in demo notebook no longer leaves behind a Recent Files entry that cannot be reopened, and a batch of invisible characters that had crept into internal log labels was cleaned out.
Three small fixes found while verifying the previous update. A harmless error that was recorded in the log file at every startup is gone. When you launch NotesXML by double-clicking the same notebook it would have reopened anyway, the app now opens it once instead of starting the open twice behind the scenes — which also means a password-protected notebook asks for its password exactly once in that situation. And a stray empty file-open event that appeared in the log on every launch-by-double-click no longer fires.
Password-protected notebooks now open when you double-click them in your file manager or drag them into the window — both used to fail with a confusing ’failed to parse’ message even though the file was fine. NotesXML now asks for your password and opens the notebook. Notebooks opened these ways also appear in Recent Files (for drag-and-drop, when NotesXML can tell where the file lives), damaged notebooks repair themselves on the way in, your saved page sort order is applied on double-click opens, and a notebook whose contents are really Markdown, CSV, Evernote, HTML or plain text opens instead of failing. Dropping several notebooks at once works exactly as before: all of them open and the first one becomes active.
PDF annotation adds three tools to the toolbar: Underline and Strikethrough for marking up PDF text (select the tool, then drag across the text), and a dedicated Arrow tool in the Shapes menu. All three export as standard PDF annotations visible in Adobe Acrobat and other readers. Also fixed: the Line tool now draws a plain line without an arrowhead — use the new Arrow tool when you want one. On scanned image pages, the text-markup tools are disabled because there is no text to select; the keyboard shortcuts now respect this too.
The Privacy Policy shown inside the app is now identical to the current edition published on the website, adding recent website-side updates that had not yet been carried into the app: the detailed description of what an optionally submitted bug report contains, the explanation that license keys are stored only as one-way hashes, the Dropbox App-folder and permission-scope disclosure, the Report AI Content documentation, and the removal of outdated location-data language. The Terms of Service, EULA, and Refund Policy shown in the app were already current.
The in-app Privacy Policy, Terms of Service, and Refund Policy now reflect that purchases made directly from iwvdigitalsolutions.com are sold through Link, a Stripe service, acting as the merchant of record. Card statements for website purchases show LINK.COM* IWV DIGITAL SOLUTIONS. The 30-day no-questions-asked refund guarantee is unchanged, and Link can additionally issue refunds within 60 days of purchase. Nothing changes for Google Play or Microsoft Store purchases.
PDFs created by the Share function for pages and notes now include the configured left and right page margins instead of running content to the paper edge. Notebook PDF sharing already had correct margins and is unchanged.
The three-dot menus for Notebooks, Pages, and Notes now use one consistent ordering (Set as Home always first, Share always next to Export) and the same Share and Export icons everywhere. Sharing a single page or note as a PDF works again; previously it could silently fail when the Professional Page Numbering print option was enabled.
Notebooks shared from Android as an NXL file are now written in the standard NotesXML file format, so the app can open them like any saved notebook. Previously a shared notebook file could be produced in an internal format the app could not read back.
Sharing a notebook as an NXL or HTML file now names the file after the notebook instead of a generic Notebook.nxl or Notebook.html, matching how PDF sharing already worked. Text and Markdown shares also show the notebook name in the heading and the share subject.
The extra empty band between the drawing toolbar and the page controls is gone on all Android devices, tablets no longer show a gap above the toolbar or a narrowed drawing area, and the bottom of the canvas now stays clear of the system navigation bar. The temporary diagnostics from the previous build have been removed.
This build adds temporary diagnostic logging to pinpoint extra spacing around the relocated drawing controls and a canvas edge issue reported on some devices. No behavior changes; the diagnostics are removed once the cause is fixed.
In the drawing editor on Android phones and tablets, the tool ribbon and the page/zoom bar have moved from the bottom of the screen to the top, just under the title. When drawing on a tablet, the hand you draw with naturally rests on the lower part of the screen — with the buttons there, it was easy to trigger tools, page changes or even Android’s own home/back gestures by accident. The bottom of the drawing area is now entirely canvas. Fullscreen drawing works exactly as before, and nothing changes on Windows or Linux.
On Android, notebooks shared as .nxl files can now be opened by tapping them in Files, Gmail, Drive and other apps — the file-type registration that made this work had been lost in an earlier update. Drawings are also no longer missing from shared output: multi-page drawings now share as one image per page instead of only the first page, every page appears in shared HTML, and shared Text and Markdown list each drawing with its title and page count (previously drawings could vanish from these formats entirely).
On Android, large notebooks reopened automatically at startup were run through a fast-but-flawed loader that could drop drawing content — notes appeared as empty placeholders until you opened the file again yourself. Startup now uses the same reliable loader as opening a file manually, which our measurements also showed is faster. Large notebooks now show a loading screen while they open instead of appearing frozen, and the app’s diagnostic log no longer loses its final entries when you quit.
If your computer has two dedicated graphics cards, large AI models are now split across both cards automatically — including the model’s working memory — so models too big for one card can still run entirely on your graphics hardware, which is much faster than falling back to the processor. When a model fits comfortably on a single card, NotesXML still uses just that card, as before. And when a model fills the available graphics memory without leaving room for its working memory, that working memory can now be kept in regular system RAM instead of forcing part of the model off the graphics cards. Computers with one graphics card are unaffected.
A hidden one-minute limit could cause converting a longer dictation to text to fail partway and start over from scratch, taking much longer than it should. NotesXML now allows the conversion the time it actually needs, based on how long you spoke and how fast your speech model runs on your device. Also, if you quit NotesXML while a dictation is still being converted to text, the recording is now saved to the recovery folder automatically instead of being lost, so you can convert it after restarting.
Previously, if converting a finished dictation to text failed — most often when a large, slow speech model needed more time than NotesXML allowed — the recording itself could be deleted along with the failed attempt. Now NotesXML gives slow models the time they actually need (scaled to how long you spoke and how fast your model runs on your device), and if converting still fails for any reason, the audio is kept safe in a recovery folder and NotesXML offers to convert it again with one tap. On Windows, longer dictations are also no longer cut short at about one minute. A backlog problem that could pile up conversion work behind a slow model mid-dictation has been removed as well, and speed measurements are no longer lost if you quit right after installing a speech model.
NotesXML now measures how quickly your chosen speech-to-text model converts speech on your own computer or phone. If you open Dictation with a model that converts slower than you can speak — which would cause long pauses between sentences and could miss words while it catches up — NotesXML tells you before you start, and offers to switch to a faster installed model with one tap, or to use Continuous Recording instead (which records everything now and converts it to text when you finish). You can always choose to dictate anyway. The measurement happens quietly in the background the first time a speech model is installed or selected, so the advice is ready before your first dictation.
When an action cannot run — for example a page or note was deleted on another device a moment earlier, or no page is open to receive a result — NotesXML now explains it in plain language and tells you what to do next, instead of showing terse internal wording like “Page not found”. A handful of status messages that duplicated an on-screen progress display have been removed, including percentage pop-ups during video conversion and media import (the operation still announces when it starts and when it finishes).
NotesXML now shows far fewer corner notifications. Messages that only repeated what you could already see on screen — a note you just reordered, an event you just deleted, a setting whose control already shows its new value, the app unlocking right in front of you — no longer appear. Cloud sync no longer announces every single file it transfers; you still get the summary when a sync finishes, and every error, warning, and confirmation of things you cannot see (files saved to disk, models installed, exports created) remains. Decorative symbols have also been removed from the remaining messages so they read cleanly with the built-in status icons.
The small notification messages that appear in the corner of the app are now easier to live with. Several messages can stack neatly instead of replacing one another mid-read, any notification can be dismissed with a click or its new close button, and error messages stay on screen longer than quick confirmations so you have time to read them. A message repeated in quick succession shows a repeat counter instead of flashing again, and screen readers now announce errors assertively. Housekeeping in the same release: a few menu entries and messages that promised features 'coming soon' have been removed or wired to the real feature — the Contact List's Import button now opens the working vCard importer instead of a placeholder message.
Android now runs the same speech-to-text engine version as Windows and Linux, bringing the same small accuracy and speed improvement to phone dictation and file transcription. All three platforms are now on one identical, verified engine build.
Android now offers the same speech-to-text model choice as the desktop app: Whisper Base (small and fast to download) and Whisper Small (more accurate, larger download) appear as cards in the setup assistant and in Settings, with install, delete, and set-active controls. Downloads are integrity-verified before use, and your selection persists across app updates. The largest desktop-only model remains exclusive to Windows and Linux. Existing installs keep working unchanged — if you do nothing, Whisper Base continues as before.
Speech-to-text descriptions now match what we measured. Larger Whisper models were described as slower, but on the updated engine the Most Accurate model (Turbo) actually transcribes recordings faster than Small on capable hardware — the real trade-off is memory and disk space, so the model list and an earlier changelog note have been corrected to say so. The setup assistant now also says plainly that the Base speech model is built into the desktop app, so speech-to-text works out of the box with no download; larger, more accurate models remain optional one-time downloads. A build safeguard was added so only intended files ship inside the speech-engine folder.
The Linux build now uses the same speech-to-text engine version as the Windows build, bringing a small accuracy and speed improvement and matching the two platforms. The engine is also now built for broad processor compatibility rather than for the machine that compiled it.
When a long recording was transcribed in sections, a few words could be lost where two sections met. The joining step now keeps everything each section produced and only removes text it can prove is already there, so nothing is dropped at a join. If it cannot be certain, it keeps the text rather than risking losing it.
In the speech-to-text model list, the highlighted card marked the recommended model rather than the one actually selected, so the list could draw your eye to a model you were not using. The highlight now marks the model in use; the star continues to mark the recommendation.
First-run setup said desktop versions bundled the speech model automatically. That was not accurate — the model is downloaded, and the larger models added recently are downloads too. The text now simply states that Whisper models support 99 languages and run fully offline once installed.
Settings, Speech-to-Text now opens the model list directly, where each model shows its size and whether it is installed, and you can install it, use it, or delete it from one screen. The old summary screen has been removed: it described the built-in model even when a different one was in use, and its Uninstall button could remove a model other than the one shown. Deleting a model now names the model and its file first, tells you if it is the one currently in use, and warns you if removing it would leave nothing able to transcribe long recordings on its own.
The Most Accurate speech-to-text model, Whisper Turbo, now transcribes recordings of any length on Windows and Linux. Long audio is processed in short, independently-decoded sections that are then joined back together, which keeps the transcript accurate from the first minute to the last — the earlier short-recordings-only limit is gone, and long files are no longer switched to a different model automatically. Whisper Small remains the recommended everyday choice for its light footprint: Turbo needs about 3 GB of memory and a 1.6 GB download, so choose it when top accuracy is worth the extra space. (Corrected in 5.150.1: on capable hardware Turbo transcribes files faster than Small, so the trade is memory and disk space, not speed.) Everything continues to run entirely on your device.
Clicking outside the Dictate window no longer discards your dictation. If you have said anything, that click is ignored and your text stays put; Cancel, the X, and the Back button now ask you to confirm before discarding. Long dictation also no longer stops silently after five minutes — there is no longer a time limit, and if the microphone stops unexpectedly NotesXML now saves what it captured instead of losing it. As an extra safeguard, dictation in progress is remembered, so if something goes wrong you will be offered the text back the next time you dictate.
On Windows and Linux you can now choose Whisper Turbo, the most accurate speech-to-text model in NotesXML, alongside Base (fastest) and Small (balanced, recommended). Turbo is best for short recordings; for longer recordings NotesXML automatically uses a model better suited to long audio, so you always get a clean transcript. Whisper Small remains the recommended default, and Turbo is an optional download. All models are verified on download and run fully offline. Full-length Turbo support for long recordings is coming in a later update.
In AI Chat's hands-free Conversational Mode, a single spoken turn can now run up to 30 seconds before the app finalizes it, instead of 12. As before, the app normally detects when you stop speaking and responds within a moment of your pause — this longer limit only matters when you speak continuously without a natural break, giving you room to finish a longer thought. Whisper Base remains the transcription engine and everything continues to run on your device.
A second speech-to-text model is now offered on Windows and Linux. Whisper Small (466 MB) transcribes more accurately than the built-in Whisper Base — on a nineteen-minute English test recording it reduced word errors by roughly a quarter — and still transcribes several times faster than real time on a typical desktop. It is recommended automatically on computers with 8 GB of memory or more; Whisper Base remains the default and nothing changes unless you choose to download and activate Small. Both models are verified against a published fingerprint on download and run fully offline. Two larger models are prepared but held back pending further testing on graphics hardware. In Settings, the Speech-to-Text buttons no longer crowd each other on narrow windows.
The 5.145.1 fix shipped the updated model catalog, but the application's internal loader still refused to read it — the file was silently rejected at startup and an older catalog was used instead, so the model list stayed hidden and a warning appeared in the log. The loader now accepts the updated catalog on Windows, Linux, and Android. This also protects Android AI model downloads, which would have stopped working with the updated catalog installed. An automated check has been added so a future catalog update cannot be silently rejected the same way.
In 5.145.0 the new speech-to-text model list did not appear in first-run setup or in Settings on Windows and Linux — both showed only the built-in Whisper Base card. The application shipped with the previous model catalog still selected, and without the updated catalog the model list correctly stays hidden. The updated catalog is now included and selected, and the model list appears as described in the 5.145.0 notes. No transcription behaviour changed in the meantime — Whisper Base continued to work throughout.
On Windows and Linux you can now choose between several Whisper speech-to-text models instead of the single built-in one. Larger models transcribe more accurately — especially in languages other than English — and run more slowly; the model recommended for your computer is starred. Find the list in first-run setup or under Settings, Speech-to-Text, Manage Models. Each download is verified against a published fingerprint before it is used, everything continues to run fully offline, and Whisper Base remains the default — if you change nothing, nothing changes. Android model selection arrives in a following update.
This update lays the foundation for choosing between several Whisper speech-to-text models instead of the single built-in one. The model catalog can now describe a ladder of speech models, each with a size, memory requirement and verified fingerprint, and NotesXML can work out which one to recommend for your hardware. Nothing changes in the app yet: transcription continues to use Whisper Base exactly as before, and no new download options appear. The visible model picker — in first-run setup and in Settings — arrives in a following update, on Windows and Linux first and then Android.
On Windows and Linux, a Table of Contents entry could point at the wrong page. If one section mentioned another by name in its text — for example “see the Sync, Import & Export page for full details” — the contents page took that mention for the section itself and listed the earlier page. Three of the twelve sections in the demo notebook were affected, and one was listed 31 pages early. NotesXML now looks for a heading rather than any mention of the title, so contents entries match the page numbers printed at the foot of each page. The same export also printed its header and footer twice on every page, one copy slightly above the other; only one is drawn now.
On Android, Professional PDF export printed the Table of Contents on the cover page, ran each section straight on from the last instead of starting a new sheet, and left a thin sliver of the previous page — the tail of a line, or a section rule — at the top of the next one. Three symptoms, one cause: this export builds its own PDF page by page, and the page-break instructions in the print layout only take effect when the operating system does the paging. NotesXML now marks those breaks in a form its own export understands. The cover page stands alone, the Table of Contents starts a new sheet, every section begins on a fresh page, and pictures, cards and figures are kept whole rather than split across a boundary — a table that genuinely runs longer than a page still breaks tidily at a row.
On Android, Professional PDF export now prints photographs. Images were being shrunk to about a single pixel by a page-fitting rule that did not apply to this kind of export, so only their captions appeared. Page breaks are also cleaner: a few lines could be drawn twice, once at the foot of a page and again at the top of the next, and a section heading could be left partly behind the footer. Each page now ends exactly where the next one begins, headings are never split, and the contents page points at the page where each heading actually appears. Page margins from your print preset are applied as before.
On Android, Professional PDF export could cut a line of text in half across a page break, and a section heading that landed on a page boundary could be split so the contents page pointed one page early. Page breaks now fall between lines and never inside a heading. The export also uses the page margins from your print preset instead of printing to the edge of the paper — the Professional preset's 1 inch margins are now applied on Android as they already were on Windows and Linux. If a notebook is too long for Professional export on your device, NotesXML now tells you how many pages it supports and offers standard printing, instead of ending without a message.
Android can now produce professionally numbered PDFs: the cover page is unnumbered, Table of Contents pages use roman numerals, body pages are numbered from 1 with correct totals on every page, and Table of Contents entries point at the exact page each section starts on — measured from the final layout rather than estimated. Because Android’s system print dialog does not allow this level of control, professional exports are generated by NotesXML itself and delivered through the Share Sheet, where you can save or send the PDF. Standard printing continues to use the system print dialog unchanged. On devices running an older version of the app, the option remains desktop-only until the app is updated.
On Android, selecting Professional Page Numbering produced PDFs with missing or repeated page numbers and up to 25 percent more pages: the technique NotesXML uses on desktop to renumber pages after the fact is not possible on Android, where the operating system produces the final PDF. Android now always prints with standard page numbering — every page numbered correctly, front matter included — and the Professional Page Numbering option in the print dialog is marked desktop-only until an Android implementation ships. Desktop printing is unchanged.
Selecting Professional Page Numbering without a Table of Contents produced a PDF with no page numbers at all: the professional mode correctly switches off the standard page numbering so the cover and contents pages stay unnumbered, but the pipeline that supplies the professional numbers was only engaged when a Table of Contents was also selected. The professional pipeline now runs whenever Professional Page Numbering is selected, with or without a Table of Contents or cover page. As a safety net, if the professional pipeline is ever unavailable, NotesXML keeps the standard numbering instead of removing numbers entirely.
When you saved a PDF, opened a file, or followed a link, the dialog or browser window could open behind NotesXML or sit as a blinking taskbar icon — most noticeably when using Print to PDF. A dialog takes on the state of the window that owns it, and NotesXML was not bringing its own window forward first, so a dialog opened from a background or minimized window stayed there. NotesXML now restores and focuses its window before presenting any dialog, and gives the hidden page-rendering window used during PDF export a proper owner so it no longer disturbs window ordering. Links and opened files also bring NotesXML forward first, which lets Windows hand the browser to the front.
Four spell-check dictionaries could never finish downloading: the transfer step handled the files as text, which silently altered their bytes, so the security check that verifies each dictionary against its known fingerprint correctly refused to install them — and for German and Polish the text handling also corrupted the dictionary content itself. Downloads are now transferred byte-for-byte, verified against the original fingerprints, and only then converted to text using the encoding each dictionary declares. All ten spell-check languages now download, verify, and install correctly on both desktop and Android.
NotesXML already verifies AI model downloads against cryptographic fingerprints pinned inside the signed app. This release extends that protection everywhere: voice downloads for Read Aloud are now verified unconditionally, every verified download records its fingerprint on your device so “Verify Installed Models” has a permanent baseline, the app cross-checks each model’s exact file size at startup, and downloads that are not listed in the verified catalog are refused rather than installed unchecked. Nothing changes in how you use AI features — this hardening runs silently in the background, keeping everything fully local and private.
After connecting Google Drive — or switching to a different Google account — the next sync could fail with an authentication error even though sign-in had just succeeded, and a second connection attempt was needed. Sync now detects when Google rejects a stored sign-in (for example after access is revoked or the app’s status changes), refreshes or re-requests authorization automatically, and retries — so one sign-in is enough. Disconnecting also now clears the last remaining stored sign-in location.
Choosing Disconnect in Google Drive sync settings appeared to work but quietly left your sign-in tokens stored on the device, so the next launch silently reconnected to the same Google account and the consent screen never reappeared. Disconnect now revokes access with Google and removes the stored tokens from every location the app can keep them, on all platforms. Reconnecting afterward always runs the full Google sign-in, letting you choose the account and review exactly what access NotesXML is requesting.
Connecting Google Drive no longer shows Google’s “unverified app” warning. NotesXML now requests access only to the files and folders it creates for sync — not your whole Drive — which is both a privacy improvement and the reason the warning disappears. Existing sync setups continue to work; your notebooks and sync folder are unchanged.
After buying Professional from Google Play or the Microsoft Store, the app unlocked your features right away — but if you opened the License page before restarting, it still showed the free-tier activation screen asking for a license key. The page now reflects your purchase immediately, showing your edition and plan without a restart. Also fixed: Microsoft Store annual subscribers were shown a “Lifetime license” label on the License page; it now correctly reads Professional (Subscription).
Managing your activated devices and enforcing the 5-device activation limit relies on our license server, and the app’s internal security policy did not include that server’s address — so those checks silently could not run. The address is now approved, and together with new server-side support, device activation, the device list under Manage Devices, and freeing device slots all work end to end. Nothing changes for existing activations.
Requesting your free cross-platform license key always failed with a network error. The app's own internal security policy, which controls which servers the app may contact, did not include our license server — so the request was blocked inside the app before it ever left your device, on both desktop and Android. The license server is now on the approved list, and key requests go through.
Lifetime purchases from the Microsoft Store or Google Play include a free license key for your other devices. Previously that key could only be requested in the moment right after purchase — if you chose Maybe Later or the request failed, there was no way to ask again. The License page now has a Get License Key for Other Devices button whenever a store Lifetime license is active, so you can request or retry the key delivery at any time.
On the Microsoft Store edition, a completed Annual or Lifetime purchase unlocked Professional immediately but the app returned to the Free tier the next time it started, because the startup check that asks Microsoft what you own could not run. That check now works. Your purchase is recorded with Microsoft either way — nothing was lost — and with this version the app recognizes it automatically at every launch. If you purchased on an earlier version, simply update: your Professional tier is restored without buying again.
On the Microsoft Store edition, choosing Annual or Lifetime in the Upgrade dialog always opened our website instead of Microsoft's own purchase window. The component that opens Microsoft's window was unintentionally disabled by a missing optional library, and the app fell back to the website without saying so. That component now runs on its own, and the app records why a fallback happened if one ever does. The website remains available as a backup path, and purchases made there continue to work exactly as before.
A link handler meant to open web addresses in your browser was mistakenly catching in-app links that use the placeholder # address, swallowing the click before the link's own action could run. This is why the startup PIN screen's Use password instead link still did nothing in 5.143.2 even after its own logic was fixed, and it also made the Upgrade to Professional banner link unresponsive. Those links now work; real web, email, and file links still open externally as before.
Two startup security fixes. First, turning off Lock on Startup (Settings > Security Settings) previously had no effect — you were asked for your PIN or password at every startup regardless. The setting is now honored: when it is off, the app opens directly. Second, on the startup PIN screen, the Use password instead link did nothing (it could even leave you on a blank locked screen); it now correctly switches to password entry, and Cancel on the password screen safely returns you to the unlock prompt instead of leaving the app stuck. Also fixed: after unlocking with your password at startup, opening a password-protected document later in the same session now correctly checks the document's own password.
The PIN field under Settings > Security Settings > PIN Protection only let you type 4 digits, even though the setup wizard and the unlock screen support 4-6 digit PINs. The field now accepts 4 to 6 digits, matching the rest of the app.
Pressing Sync Now as a Free user could do nothing at all: the sync panel button was disabled (a dead tap, especially on Android), and the Dropbox settings Sync Now could be silently skipped. Both now show the Professional upgrade message so you always get a response. If you have already seen the message once this session, automatic syncs will not repeat it.
On the Free tier with cloud sync configured, the Professional Feature upgrade message could appear again right after you dismissed it, because two automatic sync attempts each showed their own copy. The message now appears at most once per session. Choosing Not Now really means not now.
If you set a 5- or 6-digit startup PIN, the unlock screen previously showed only four boxes and could never accept the full PIN, locking you out of the app. The unlock screen now shows the correct number of boxes for your PIN, plus an Unlock button, so 4-, 5-, and 6-digit PINs all work. If you were affected, simply enter your real PIN on this version to get back in — no data is lost.
On the Free tier, the Professional Feature upgrade prompt (for example, the one shown when turning on Cloud Sync) could occasionally appear with its Upgrade, Not Now, and tap-outside actions all unresponsive, and the Esc key did nothing. This happened when the prompt was shown more than once in a single session. The prompt now clears any earlier copy before appearing, every button responds reliably, and pressing Esc (or the Back button on Android) closes it.
If a Professional-tier AI model was selected while licensed, it previously stayed selected and working after the license was deactivated or expired. Now, when a license ends, the app automatically switches to the best free model (your downloaded model files stay on disk, and re-activating a license restores access instantly). The AI engine also independently verifies the selected model's tier before every session.
On Free-tier installations (including after a trial ends), AI Chat could still be opened from the AI toolbar button — the license check covered one way of opening the chat but not the toolbar route. All ways of opening AI Chat now show the upgrade prompt on the Free tier. Professional licenses are unaffected.
After the Beta label was retired in 5.142.91, eight of the app's built-in self-test checks still expected version numbers to start with 'Beta' and reported failures on a healthy installation. Those checks now accept the current version format (and still recognize older Beta-labeled builds). No app functionality was affected — only the test suite itself.
The 'Purchase a license' link on the license activation screen pointed to a web page that does not exist, showing a 'page not found' error. It now opens the current pricing page. Every website link in the app was also checked against the live site as part of this fix, and an automated check now prevents broken links from shipping in future releases.
On the Microsoft Store edition, the Annual and Lifetime buttons in the Upgrade dialog now open Microsoft's own purchase window — pay with your Microsoft account, no license key needed, and your purchase follows your Microsoft account. A new 'More purchase options' link offers direct website purchase (including academic pricing) as an alternative. Also, NotesXML has graduated: the version label no longer carries the Beta prefix.
Clicking Annual or Lifetime in the Upgrade dialog on the Microsoft Store edition opened the website at an old address that no longer exists, showing a 'page not found' error instead of purchase options. The buttons now open the current pricing page. The academic verification service was pointed at an outdated address with the same problem and has also been corrected.
Misspelled words were underlined in red, but right-clicking them offered no way to fix them — the context menu showed only editing commands. Right-clicking a misspelled word now lists up to five suggested corrections (click one to replace the word) plus an 'Add to Dictionary' option, followed by the usual editing commands. On Android, the system's own text-selection toolbar already provides spelling suggestions and is unchanged.
On Linux systems where audio plays through a TV or monitor (HDMI or DisplayPort), the first word of spoken responses could be partially cut off — these audio outputs take up to a second to wake after sitting idle, longer than the built-in silence padding. NotesXML now keeps the audio output awake during speech sessions, so every response starts cleanly. This is the desktop counterpart of the Android fix in 5.142.84.
When an AI response contained only numbers or a plain equation (like a list of digits or '2 + 2 = 4'), the text-to-speech skipped it entirely — a filter meant to suppress stray list markers was discarding all content without letters. Numbers and equations are now spoken, and simple math reads naturally ('2 plus 2 equals 4', '15 percent'). Stray list markers between sentences remain correctly silent.
When the AI spoke a response or read a note aloud on Android, the first word was clipped or lost entirely — the device's audio output takes a moment to open, and the beginning of the speech was swallowed while that happened. NotesXML now holds the audio output open during speech sessions, so every response and every reading starts cleanly from the first word.
When importing a large file (such as a big CSV), the 'Opening File Dialog' helper message could remain on screen through the entire import — long after you had already chosen the file — and disappear only when the import finished. The message was actually being dismissed on time, but the import work started so quickly afterward that the screen never got a chance to redraw. The app now guarantees one screen update between closing the dialog and starting the import, so the message disappears the moment you pick a file and only the import progress indicator remains.
The previous fix connected the video player to Android's secure storage, but the app's built-in content security rules — which had no rule for media — were still blocking the video before it could load. A media rule now explicitly allows the app's own stored videos. Desktop playback is unaffected. The video format itself was verified fine all along.
Videos saved to your notebook's Media/Video folder on Android imported correctly but showed 'Video could not be played' when opened. Android's storage security blocks the direct file path the player was given, even though the app saved the file through the proper channel. The player now opens videos through the same secure storage permission used to save them, so externally stored videos play normally.
A cluster of controls were wired to functions that never existed, so they failed without any message: the Ctrl+Shift+S Save As and Ctrl+O Open shortcuts in the mobile drawer, CSV and HTML imports from the drawer (the file was read and then discarded), all file imports launched from the search palette (Evernote, Markdown, HTML, text, XML), and the calendar Contact List, which always showed 'No contacts yet' even after you created contacts and threw errors when you typed in its search box. All of these now call the real functions: the shortcuts save and open, the imports import, and the Contact List shows, searches, sorts, and filters your contacts. JSON import, which was never actually supported, now says so instead of doing nothing.
Importing a video and choosing 'Save Externally' (or importing any video over 100 MB, which requires external storage) always failed on Android with 'External storage not available on this platform'. Android can now save videos to your notebook's Media/Video folder just like desktop. Videos are transferred in small pieces, so even very large files import without exhausting memory.
When printing with Professional Page Numbering and Section Headers turned off, the printed table of contents showed no page numbers at all (body pages numbered correctly). The hidden markers the print system uses to locate each section were being dropped from the PDF's text layer by the rendering engine, so the table of contents could not find where sections started. The markers now survive, and every table of contents entry shows its correct page number.
If your antivirus quarantined part of the AI engine (this has happened with Avast), NotesXML used to retry pointlessly and then show a generic startup error. It now detects the missing file immediately and tells you exactly what happened and how to fix it: restore the file from your antivirus quarantine and add an exclusion for the NotesXML folder, or reinstall NotesXML.
On Windows and Linux, calendar reminders can now also appear as system notifications (in addition to the in-app alert window added in the previous update). Clicking the notification brings NotesXML to the front. You can turn this on or off under Settings > General > "System notifications for reminders" — it is on by default. Note that your operating system's own notification settings (such as Windows Focus Assist) still apply.
In timezones behind UTC (all of the Americas), calendar items could display — and in some views, be filed — one day early: an event note dated July 16 showed "July 15," tasks near a month boundary could appear under the wrong month, overdue calculations could be a day off, and a today event could vanish from Upcoming Events. All fixed. Also: reminders for timed events and tasks used to appear only as a brief 10-second popup message, easy to miss — every reminder now opens a proper alert window that stays until you dismiss it, and a reminder time you set yourself is now honored for timed items too.
If a task reminder and an event reminder fired at the same minute, the second one silently replaced the first — you could miss a meeting alert without ever seeing it. Both reminders now appear together and are dismissed separately. Also fixed: the password strength meter in Settings > Security never responded while you typed (it was accidentally wired to a different, hidden meter); it now works, and resets properly when you cancel.
Two components were silently missing from Windows packages: the built-in English spell-check dictionary (spell check needed an internet connection the first time you used it) and a small helper that instantly frees your graphics card's memory if NotesXML is force-closed or crashes during AI use (without it, the memory was only reclaimed the next time you started NotesXML). Both are now built, verified, and included in every Windows release.
When AI features were switched off, searching the command palette for "Settings: AI" found nothing — the one entry that leads to the switch for turning AI back on was hidden along with the AI features themselves. The palette now always lists Settings: AI (except in editions built without AI), so both routes to re-enabling AI — the settings menu and the palette — stay open.
If you switched off "Enable AI Features," the AI Model settings page refused to open afterward — which meant the switch to turn AI back on was unreachable, even after restarting. The settings page now always opens (the switch lives there), so disabling AI is a choice you can reverse at any time. Also removed a duplicate, non-working "Enable Punctuation Enhancement" toggle from the model manager that could silently affect the main AI switch; punctuation enhancement follows the main AI Features switch.
Since late June, the AI engine could silently run on your computer's processor even when a capable graphics card was available, making AI responses much slower than they should be — on Windows and Linux alike. The way NotesXML asked the AI engine which graphics cards exist had stopped working after an engine update. Hardware detection now asks directly and reliably, works with any number of graphics cards, and re-tests automatically. If AI felt slow recently, this update should restore full speed; the first launch will re-check your hardware once.
Added deeper hardware-detection notes to the on-device technical log introduced in 5.142.66 (graphics-driver configuration seen by the AI engine at startup — never your notes or words), so the remaining cases where AI does not use a capable graphics card can be diagnosed from a single report. This build also consolidates the 5.142.66 microphone-detection fix and technical log, which some 5.142.67 build trees were missing.
Entering your license key by pasting it (instead of typing it manually) could scramble the key if its first character after the type prefix happened to be a letter — for example turning PRO-ABCDE-... into the invalid PROA-BCDE-... — causing a valid, purchased key to be rejected as "Invalid key format." The key entry field now recognizes the PRO, ACAD, BETA, VIP, and NXML prefixes directly, so pasted keys are grouped correctly no matter what their first character is. If you previously had trouble activating a key, try again — it will now paste in correctly.
The Microsoft Store edition could fail to find your microphone even after permission was granted, because of how it asked Windows for the device list. It now uses a more reliable method. NotesXML also keeps two small technical notes on your device (about microphone detection and AI hardware selection, never your notes or words) so problems like this can be diagnosed from a single report.
In the Microsoft Store edition, dictation and speaking to AI Chat failed with a microphone permission error, even when your microphone was working everywhere else. The Store package was not asking Windows for microphone access, so Windows refused it and NotesXML could not appear in your microphone privacy settings. It now requests access correctly. Windows will ask your permission the first time you dictate.
On Windows and Linux, NotesXML was accidentally hiding your graphics card from its own AI engine, so AI ran on the processor instead — much more slowly — even on machines with a powerful GPU. Nothing reported an error. This is fixed, and NotesXML will re-check your hardware once on the next launch and start using the GPU if you have one. If your AI has felt slow, this is likely why.
On Android, the components NotesXML uses to talk to Google Play for Professional purchases were updated to Google's newest billing library, as Google requires for all app updates later this year. Purchases, subscriptions, and restoring an existing purchase work exactly as before. If a product can't be offered on your device, NotesXML can now record the specific reason, which helps us fix store problems faster.
Every part of the Report AI Content dialog now follows your chosen theme, including the category dropdown and the additional-detail text box, which previously stayed white on dark themes. Colors come from the same theme system used by the theme picker, so text always stays readable on every theme.
The new Report AI Content dialog now follows your chosen theme like the rest of NotesXML, and its buttons use the same styling as other AI dialogs instead of plain system buttons. The report text itself now also includes the support email address at the top, so if you use Copy Report instead of Send by Email, the copied text tells you where to send it.
Internal build tooling only: the pre-packaging check now understands how the bundled AI engines are packaged today (a small launcher alongside an engine library) instead of expecting older single-file layouts, and now also verifies the speech engine's core library is present. No change to the application itself.
The AI, dictation, and read-aloud engines previously required a Microsoft system component (the Visual C++ Redistributable) that most Windows PCs already have — but on a fresh Windows installation without it, those features would silently fail to start. NotesXML now ships the three required Microsoft runtime files inside the application itself, so the AI engines run on any Windows machine with nothing extra to install. Nothing changes on machines that already worked, and Linux and Android are unaffected.
You can now report AI-generated output that you find harmful, offensive, misleading, or otherwise inappropriate. A Report button appears on every AI Chat response and on every AI action result (Summarize, Translate, Polish, and the rest), and there is a permanent 'Report AI Content' entry under Help & Info. Reporting does not change how NotesXML handles your data: the AI still runs entirely on your device, and nothing is transmitted automatically. Choosing Report builds a plain-text report on your machine, shows you the exact contents, and hands it to your own email app — it is not sent until you send it. The AI's response is included by default, since that is what you are reporting; the prompt or source note that produced it is left out unless you tick the box, because that may contain your private notes.
Completes the previous update: an internal recording detail meant the live dictation text still could not start on Android. Words now appear while you speak, refreshing every few seconds depending on your device.
On Android, the dictation window previously stayed empty while you talked, even though your speech was being captured correctly — the words only appeared in the finished note. The live view now works on Android just as it does on desktop: text appears while you speak and firms up as you pause. On slower phones the live text may update every few seconds; your full dictation is always captured regardless.
Two fixes for extended dictation sessions. First, the live text no longer freezes after roughly a minute of continuous speaking without a pause — the speech engine was re-reading the entire recording for every live update, taking longer and longer until it timed out; it now reads only the most recent stretch, so updates stay fast no matter how long you talk. Second, the view now reliably follows the newest words: previously a single large text update could silently stop the automatic scrolling for the rest of the session. Scrolling up to review still parks the view; returning to the bottom resumes following.
The quick-insert pins next to the + button (which notes you pinned, and the order you pinned them in) are now saved with your other settings and restored exactly on the next launch. Previously they could silently revert to the defaults after closing and re-opening the app. If you had a pin set that kept getting lost, it will start sticking from the first change you make after this update.
During long dictations the live text view now scrolls automatically so the newest words are always visible. Previously, once the view filled up, new text was added out of sight below and dictation appeared to have stopped. If you scroll up to re-read something while dictating, the view stays where you put it and resumes following once you scroll back down.
A small correction to the reorganized Insert menu: Video note loads an existing video file from your device, so it now lives in the Attach group with Image and Gallery rather than under Capture. Capture is now strictly for creating new recordings and photos.
Two refinements. First, the Insert (+) menu groups now follow what each action really does: Audio moved into Capture beside Continuous recording (recording is capturing); Image and Gallery moved into Attach (they insert existing files — the camera captures new ones); Table moved into Notes beside the lists (a table is content you write); and the last group is now called Organization — tasks, events, calendars, contacts, indexes and their management tools. Your pinned shortcuts are unchanged. Second, on Android, dictation and transcription now record through the voice-recognition audio source that speech engines are designed for, which can improve accuracy on many devices; the automatic pause detection may recalibrate itself over your first session or two.
The floating recording bar shown during Continuous recording now displays the same live level meter as the Audio note recorder and Dictation — the color-zoned bar with the peak line — driven by the actual microphone signal. Previously the bar showed an animated decoration that looked like a waveform but was not connected to the sound. If you minimize the bar, the meter pauses to save battery on long recordings and resumes when you expand it. On a small number of configurations without a live signal source the decoration remains.
The Insert (+) menu is now deliberately organized. Dictate moved up into the Notes section, right beside Text — it creates a text note, and now it reads that way. Within every section the most-used items come first and related items sit together: the list types are grouped, image capture options are side by side, tasks and events and contacts each keep their pairs together, and housekeeping actions like Manage calendars sit at the bottom. The Attach section (attachments and links) now appears above Structure. Your pinned shortcuts are untouched.
Dictation now shows the same professional level meter as the Audio note recorder: the color-zoned bar (yellow when too quiet, green when good, red when too loud) with the peak-hold line and level markers. One meter, one meaning, on every surface that listens. Under the hood the meter drawing is now a single shared implementation used by the Audio note recorder on both desktop and Android and by Dictation, so the three can never drift apart again.
Dictation (added in Beta 5.142.44) now works the way you speak: a pause simply finishes the sentence — the text moves into the note preview and listening continues, so you can dictate a whole paragraph hands-free. Press Stop & Insert when you are done to save everything as one text note; Cancel still discards. A live level bar shows the microphone hearing you (blue while you speak). If you stop talking entirely, dictation ends itself after a minute — saving what you said, or closing quietly if there was nothing — so the microphone is never left open. On rare audio setups without live pause detection, the dialog tells you to use Stop & Insert instead.
A new Dictate entry in the insert (+) menu lets you speak instead of type. Your words appear live as you talk, and when you stop (or pause for a couple of seconds) the text is saved as a new text note. Dictation is free on every tier with no time limit — it is an input method, like the keyboard. Only the text is kept; the audio is never stored. Whisper already adds punctuation and capitalization as you speak; for further polish, the free Grammar Polish and Clean Up Formatting actions are one tap away in the note AI menu. Dictation is different from transcription: transcription converts saved audio and video recordings into text and remains a Professional feature (the Free tier transcribes up to 30 seconds). Requires the one-time Whisper speech model download (offered on first use).
When a recording is longer than the Free tier's 30-second transcription limit, the upgrade dialog now offers a third choice: Transcribe first 30 seconds (Free). Choosing it transcribes just the opening of your recording so you can see the transcription quality on your own audio; the result ends with a note showing where it stopped and how to unlock full-length transcription. Works for audio and video notes, the note menu Transcribe action, and the Smart Recorder. On Android the preview is offered for recordings up to 15 minutes long (a memory safeguard); on desktop there is no source-length cap. Professional and trial users are unaffected.
Closed a gap where recording with the Smart Recorder in Record & Transcribe mode, or using its in-recorder Transcribe button, could transcribe recordings longer than 30 seconds on the Free tier. The 30-second Free-tier limit introduced in Beta 5.142.40 now applies on these paths too. Recordings over the limit are always saved as audio notes — nothing is lost — and an upgrade prompt explains the limit. Professional and trial users are unaffected (up to 4 hours). Voice input and dictation remain unlimited on all tiers.
Fixed the Upgrade to Professional dialog (and five related license dialogs: Limit Reached, Trial Expired, Cross-Platform Offer, Subscription Lapsed, and Build Expired) appearing see-through, with the page showing through the dialog and its buttons hard to read. The cause was a malformed style rule that made the dialog background invalid, so the browser dropped it entirely. These dialogs also gained a visible X close button, close when you click outside them, and close with the ESC key (or the Back gesture on Android). The dialogs now correctly follow your light or dark theme.
The Free tier can now transcribe audio and video recordings up to 30 seconds long — enough to try the offline transcription feature before upgrading. Longer recordings show an upgrade prompt and are kept as audio or video notes; nothing is lost. The Record & Transcribe mode can be selected on the Free tier (a notice explains the 30-second limit). Professional and trial users are unchanged (up to 4 hours). Also, transcribing handwriting in drawing notes and running OCR on image galleries now correctly require a Professional license, matching OCR everywhere else in the app.
On the Free tier, AI menu items that require a Professional license now show a small crown badge, and the page-level and notebook-level AI menus show a single clear upgrade line instead of a list of items that would each refuse when tapped. The PDF text extraction dialog also marks its Professional post-processing options (Auto Title Pages, Generate Notebook Title, Executive Summary) instead of showing checkboxes that silently did nothing. Nothing changes for Professional or trial users.
Fixed an issue where several Professional features did not check your license and were available on the Free tier: PDF annotation, some cloud sync entry points, one of the two transcription paths, background audio recording, export without branding, and editing AI prompts. In addition, the Professional AI actions at the page and notebook level (such as Summarize Page, Executive Summary, and Table of Contents), the note-level Translate, Extract Action Items, and Custom Prompt actions, and the OCR, Analyze Diagram, and Scan Receipt vision actions now correctly ask for a Professional license on the Free tier. Nothing changes if you have a Professional license or an active trial — every new install still includes 14 days of full Professional access. Free features (including Describe Image, Auto Title, Polish, and Summarize on a single note) remain free.
After you saved a notebook, a later save (including automatic saves) could incorrectly warn that the notebook was modified outside the app, even though the only change was your own save. The app now correctly recognizes its own saves on every save path. Also, automatic background saves will no longer pop up the "changed on disk" question while you are away — if a real conflict is detected, the automatic save quietly waits and your next manual save will ask you how to resolve it.
On a page with many notes, or a very long note, the scrollbar handle on the right edge of the content area could shrink to a sliver and was nearly invisible on dark themes until you hovered over it. The handle now has a guaranteed minimum size no matter how long the content is, and its color adapts to your theme so it stays visible on light and dark themes alike. The scrollbar is also wider than the old default, making it easier to click and drag. It brightens when you hover or drag it, and it grows and shrinks together with the Interface Scale setting like the rest of the interface. Applies on desktop and Android.
Fixed an issue where the Gemma 4 E4B model could garble dates, numbers, and technical values when answering questions about long documents — for example, quoting a date or figure slightly differently from what the document actually says, or mixing foreign characters into its answers. The cause was an anti-repetition setting (added earlier to stop runaway output during scanned-document OCR) that also discouraged the model from repeating facts exactly as written in the source. The setting is now scoped so runaway-output protection is preserved while facts from long documents are reproduced faithfully. Other models are unaffected.
Fixed a deadlock during AI actions on documents containing hidden content: the 'Security review — Hidden content detected' dialog could appear behind the AI progress window, making its buttons unreachable and leaving the action stuck. The security dialog now always appears on top; choose an option and the action continues normally.
Data safety (important): fixed two serious issues where note content could be lost when saving. Files attached as notes (any file type) saved in one session lost their payload and showed 0 bytes after reopening; and images added as image notes could disappear after reload. Both are fully fixed — new saves are safe. Items affected before this update cannot be recovered from saved notebooks or their backups; please re-attach those files and images from the originals.
AI now reads attached documents everywhere. Page and notebook AI actions (Summarize, Briefing, Key Themes, Conflict Analysis, and more) now include the content of attachments and file notes — Word, Excel, PowerPoint, PDF, and text files — instead of skipping them. Notebook Briefing, Key Themes, and Conflict Analysis use a new hierarchical analysis that processes each page (including its attachments) before producing the final result. Large documents of any size can be summarized: AI Chat offers a “Summarize as Note” option for oversized attachments, and AI actions on file notes process big documents in chunks with progress shown.
New formats: OpenDocument (.odt, .ods, .odp), EPUB e-books, and RTF are now supported by AI features and text extraction. Also fixed: a required ZIP-reading component referenced by the app had never been included in the package — shipping it restores PowerPoint text extraction and Office-document search indexing, and enables the new formats.
New setting: AI Settings → “AI Document Text Limit” controls how much text AI features read from a single document (default 2 MB — ten times the previous limit; adjustable 0.1–10 MB). Documents over the limit are truncated with a clear notice.
Also in this release: the AI Chat speaker button no longer disappears when text-to-speech is briefly unavailable (it now shows an unavailable state and recovers automatically); AI-generated notes from page and notebook actions display formatted text instead of raw markup; text file contents attached to notes are decoded correctly in AI summaries; and saved web pages no longer overflow the AI context window during summarization.
Fixed AI actions (Summarize, Generate Title, Translate, Extract Action Items) on file notes containing Word, Excel, PowerPoint, OpenDocument, or EPUB documents: the AI previously received an internal placeholder instead of the document text and typically replied asking for the document. These actions — and the notebook-level summaries that include such notes — now extract and use the real document content, at any size (large documents are processed in chunks).
Fixed the recommended 'Summarize as Note' option in the dialog that appears when a document attached in AI Chat is too large for the AI context window. Selecting it previously did nothing useful; it now summarizes the full document in chunks (any size), shows progress, and offers to save the summary as a note on your current page.
Fixed the error 'extraction library not available' when attaching OpenDocument files in AI Chat. A required component (the ZIP reading library) was referenced by the app but had never been included in the package — this also silently prevented PowerPoint text extraction and Office-document search indexing. The library now ships with the app, enabling PowerPoint (.pptx), OpenDocument (.odt/.ods/.odp), and EPUB reading everywhere AI features extract document text.
The notebook-level Briefing Document, Key Themes, and Conflict Analysis prompts now analyze the notebook page by page — including attachments — before producing the final result, giving more thorough and better-organized output on large notebooks (per-page findings are included below the final analysis). Conflict Analysis extracts each page's specific claims, dates, and figures first so genuine contradictions are preserved for comparison. The page-level Summarize prompt now summarizes each attachment individually before combining. Also new: files attached in AI Chat now include EPUB books (with chapters read in order) and RTF documents (read as clean text instead of raw formatting codes); EPUB files are also readable by page and notebook AI prompts and file notes.
AI features now extract text from OpenDocument files: Writer (.odt), Calc (.ods), and Impress (.odp). This applies everywhere document text is read — page and notebook AI prompts reviewing attachments, AI actions on file notes, and files attached in AI Chat. Very large documents are trimmed to the same generous per-file limit as Word, Excel, and PowerPoint files. Also fixed: PowerPoint and OpenDocument files can now be chosen in the AI Chat attach dialog (previously PowerPoint files could only be dragged in).
Fixed an issue where an image added to a page (for example by dragging an image file in, or from an AI Chat image) displayed correctly during the session but was empty after saving and reopening the notebook. The image data is now stored correctly. Images affected before this fix cannot be restored from the notebook — please re-add them from the original files.
Fixed a serious issue where a file attached as a note (for example by dragging a document into a page) lost its entire content when the notebook was saved: after reopening, the file showed 0 bytes and could not be opened. The file data is now stored correctly and survives save, close, and reopen. IMPORTANT: files affected before this fix cannot be restored from the notebook — please re-attach them from the original files. PDF notes were never affected.
Fixed an issue where AI summaries involving certain text files contained scrambled, code-like text. Files stored in a particular internal format were being passed to the AI without being decoded first; they are now converted to readable text before any AI processing.
The notebook-level AI summary now follows the full hierarchy: each page's attachments are summarized individually (short ones included as-is, unreadable ones listed by name), each page is then summarized, and a final overall summary is produced. Pages containing only attachments (no notes) are no longer skipped. Also fixed: summarizing a page with a saved web page (.html) attachment could fail with a context error — hidden script and style code is now excluded from the extracted text, and very large content is split safely instead of being rejected.
Fixed an issue where the AI Chat text-to-speech speaker button could be invisible until audio had been produced elsewhere in the app. The button is now always shown: when text-to-speech is not yet available (for example, right after launch while the speech engine is still starting), the speaker appears with a red prohibition mark and an explanatory tooltip instead of being hidden. It recovers automatically within a few seconds once the engine is ready, or immediately when clicked.
AI prompts run at the page or notebook level (Summarize, Custom Prompt, and the other built-in prompts) now include the content of file attachments, not just notes. PDF documents (with a text layer), Word, Excel and PowerPoint files (.docx/.xlsx/.pptx), and text-based files (such as .txt, .md, .csv, .json, source code and logs) are read, with very large files trimmed to a generous per-file limit. Attachments that cannot be read — images, legacy Office formats, audio/video, scanned PDFs — are listed by name in the AI's input as not included, so the response is honest about what it covered. Also fixed: AI actions on text-based file notes previously received the file's raw encoded form instead of its readable text.
Fixed an issue where saving an AI response as a note produced unformatted text (literal ## and ** symbols) even though the response preview showed proper headings, bold text, and lists. Affected: Custom Prompt at the page and notebook level, the built-in page and notebook AI prompts, and Translate Page. The response is now converted to formatted content when the note is created, matching how AI Chat and note-level actions already behaved.
Fixed an issue where downloading the French spell-check dictionary failed. The French dictionary is published in a different location than the other languages, and the app was looking in the wrong place; it now downloads correctly, pinned to a fixed version and verified against a built-in SHA-256 fingerprint like the other languages. All ten downloadable dictionaries (English US/UK, French, Spanish, German, Italian, Portuguese Brazil, Dutch, Russian, Polish) are now integrity-checked. Also fixed: on first use of an AI feature, the one-time AI content disclaimer could appear underneath the processing spinner, leaving the app stuck; the disclaimer now always appears on top and the spinner waits until you continue.
Follows the 5.142.17 groundwork. Spell-check dictionaries that are downloaded on demand are now pinned to a fixed, verified version and checked against a built-in SHA-256 fingerprint before they are installed, for English (US/UK), Spanish, German, Italian, Portuguese (Brazil), Dutch, Russian and Polish. A download that does not match is rejected. (French is temporarily excluded from this check because its dictionary is currently unavailable from the upstream source; that is being addressed separately and no French behavior changes here.)
Internal hardening from a pre-release review, with no change to how you use the app. Text produced by AI features now preserves legitimate right-to-left and left-to-right direction marks (used in Arabic, Hebrew and mixed-direction text) while still removing hidden or invisible characters, and every AI feature now shares a single text-cleanup step. Feature-access checks now fail safely to the Free tier in the rare case the licensing component has not finished loading, instead of momentarily allowing access. Spell-check dictionaries that are downloaded on demand can now be pinned to a fixed version and integrity-checked (SHA-256) before they are installed. A new build-time check prevents a paid-distribution build from ever being packaged with licensing turned off.
Internal test-suite maintenance with no change to the app you use. Two checks in the built-in self-test were corrected: the catalog schema-version check now reads the app's own list of supported schema versions (it had a stale hardcoded copy that started failing after the model catalog moved to schema 2.2), and the checklist-checkbox handler check now actually verifies current behavior (it had been silently passing without testing anything). No app behavior changes.
Two refinements to the AI Models screen. The model tier descriptions (Basic, Standard, Full, Premium, Power) now describe what each tier is best suited for — speed and output quality — instead of implying that lower tiers cannot run certain features. Every model can run every AI feature; the tiers differ in speed and quality. Also, the Verify Installed Models button (Windows and Linux) now sits just below the catalog selector, so its results list no longer appears wedged between the download buttons.
Internal cleanup with no user-facing change: the one-time AI disclaimer is now enforced from a single place in the code (the point where any AI feature starts the model) instead of being repeated across individual features. This follows the 5.142.13 change that made that single place present the disclaimer rather than an error. The disclaimer behaves exactly as before.
Fixed an issue where choosing an AI action before you had acknowledged the one-time AI disclaimer (that AI results can be inaccurate and should be reviewed) could show an error message. NotesXML now presents the disclaimer for you to accept, then continues with your action; if you decline, the action is simply cancelled with no error. This now applies consistently across every AI feature. No change if you already accepted the disclaimer.
The AI Models screen now has a clear way to return to the built-in, verified model list. If you had switched to a custom or imported model catalog, its models show as Unverified and are not integrity-checked; you can now choose "Use verified default catalog" (from the Switch catalog menu, or directly from the Verify Installed Models message) to return to the built-in list, which is part of the signed app and cannot be altered on your device. Switching back re-enables integrity checking. Your custom model list is hidden, not deleted — it stays available under "My Custom Catalog (your edits)" and your selected model is preserved where possible.
When you download a Read Aloud voice, NotesXML now verifies it against a fingerprint (SHA-256) built into the app: the downloaded voice file is matched to a known-good value that ships inside the signed application, so a corrupted or altered download is rejected instead of installed. Voice downloads are also now pinned to a fixed, unchanging version of each file. This is the same protection already applied to AI models. No action is needed on your part; any voices you have already installed are unaffected.
Resolved an issue where Verify Installed Models reported an error for larger models (roughly 2 GB and up) on Windows and Linux, while smaller models and vision packs verified correctly. The check was reading each file into memory all at once, which failed on large files; it now reads them in a streamed fashion, so every installed model verifies regardless of size. Your model files and their fingerprints were never the problem — only the check itself. Re-run Verify Installed Models to see all models pass.
Resolved an issue where AI model verification did not take effect on Windows and Linux: the desktop app was not accepting the updated model list format, so verification stayed off. With this fix the app reads the verified model list correctly, so downloaded models are checked against their built-in fingerprints and the Verify Installed Models button works as intended. If you installed models on the previous version, re-open the AI Models screen; no re-download is required.
Downloaded AI models are now checked against a fingerprint built into the app: each model file is matched to a known-good value that ships inside the signed application, so a corrupted or altered download is rejected instead of installed. The Verify Installed Models button (Windows and Linux) now confirms your already-installed models pass this check, file by file. Model downloads also now come from a fixed, unchanging version of each file. No action is needed on your part; your installed models are unaffected.
Fixed two settings on the newly added Ministral 3 14B model so it matches the other Ministral models: its images are now prepared at the correct 1024-pixel size (was 896), and its repetition setting now uses the standard value instead of being left unset. This resolves two internal self-test warnings. Model quality was already good; this brings the model fully in line with the rest of the Ministral family.
The AI Models screen now offers ten downloadable models instead of eight. New additions: Gemma 4 E2B, a compact vision-capable model suited to phones and smaller PCs, and Ministral 3 14B, a larger vision model for desktops (it needs about 11 GB of RAM, so it is offered on Windows and Linux only). The model list is now ordered from smallest to largest, so you can pick one that fits your device at a glance. Your free-tier models and default selection are unchanged.
Every AI model that NotesXML downloads is now verified before it is used. The app computes a fingerprint (SHA-256) of the downloaded file and compares it against a value built into the signed app; any file that does not match is rejected and removed. Download links are pinned to fixed repository versions, so the models offered by default are locked at release and were tested with this exact version. The in-app 'Check for Model Updates' and 'Load Custom Model' options have been removed — model availability now ships with app updates. Advanced users can still load their own model catalog, but those catalogs now show a one-time notice and mark their models as Unverified, because they are not tested with this version or fingerprint-checked. You can re-check your installed models at any time using the new 'Verify Installed Models' button on the AI Models screen (Windows and Linux).
The AI Models screen now makes clear that downloading a model adds a data pack — no software is installed on your device. (Windows only: the Start-menu and taskbar icons now show the NotesXML logo, correcting a generic placeholder.)
On Microsoft Store builds, the upgrade and trial-ended screens now show both purchase options: an annual subscription at $29.99/yr and the one-time $49.99 lifetime license, matching the pricing on our website. Nothing changed about what you own or pay — this only corrects the price summary shown inside the app.
A build-system correction with no changes to the app itself.
Tables now support selecting whole rows, whole columns, or groups of them. On desktop, hover over a table and click the slim handles at the left of a row or the top of a column (Shift+click extends the selection). On Android, long-press a cell in the table editor and choose Row or Column; long-press another cell to extend. A floating toolbar appears with Copy, Cut, Paste, Insert, Clear contents, Copy as CSV, and Delete (with Undo). Paste replaces the selection and understands data copied from Excel, Sheets, and LibreOffice. Pasting a block of cells into a single cell now offers to split it across cells. Also fixed: pasting multi-line text into a table cell no longer creates a new note by mistake.
When a sync wants to delete more files than your safety limit, NotesXML now shows the count with examples and asks for confirmation instead of silently skipping the batch and reporting a sync error. The new Safe Delete Threshold setting (Sync Preferences → Deletion Safety, default 10) controls when the question appears. Also fixed: deletions recorded while offline now reliably propagate to Google Drive.
Sync's stall detector could not see folder-listing activity, so accounts with hundreds of folders were incorrectly aborted mid-listing with a “Sync stalled” error on every attempt. Listing progress now counts as progress; genuinely hung connections still time out as before.
If a file that was deleted in the cloud is already gone from your computer, sync now records it as complete instead of reporting an error and retrying on every sync. Existing stuck entries clear themselves automatically on the next sync.
Catalog documentation backfilled and internal copies brought back into exact sync — no model, default, or pricing-tier changes. Purely behind the scenes: it keeps the AI model list dependable across app updates.
Insert and delete row/column in table notes now act on the cell you clicked — previously, opening the menu or ribbon made the app forget your selection on desktop, so Delete Column could remove the wrong (last) column. If no cell is selected, the app now says so instead of silently acting on the last row or column. New: Copy Row, Cut Row, Copy Column, and Cut Column on the table ribbon (and the Android table editor) — rows copy as tab-separated text, columns as a list with the header, ready to paste anywhere.
NotesXML now checks at startup that the bundled AI engine (llama-server) is the exact build this version was tested with, and warns you if an installation problem left an older engine in place — the cause of the vision failures diagnosed on 2026-07-06. Build tooling gained matching safeguards so a mismatched engine or model catalog can no longer ship silently.
Running a test script (Settings > Run Test Script) now tells the script its own file location. The Test Orchestrator v2.3 uses this to scan the folder it was launched from automatically — no more clicking through the directory picker to point it at the test folder it is already sitting in. If the folder holds no other test scripts, or on Android, the familiar picker appears as before.
Behind-the-scenes fix: code blocks, indented text, and math formulas in rendered markdown are no longer wrapped inside an invalid paragraph element. Browsers were silently correcting this, so nothing changes visually — the rendered document structure is simply valid now, which also lets automated tests check it reliably.
The third and final route by which a list of dates could turn into a code block: pasting. Copies from terminals, file managers, and code-styled pages carry hidden monospace markup on the clipboard, and pasting that into a note created a code block directly — no importing or AI involved, which is why the two previous fixes didn't catch it. Pasted monospace content now arrives as normal text, one line per line. Deliberate code is unaffected: code blocks copied between your own notes keep their structure, typing ``` still creates one, and the Code button still converts selected text into one. Speaking of which — the Code button previously could not take you back OUT of a code block (it was blocked by the same rule that blocks bold there), so an accidental code block was a trap. Pressing Code inside a code block now converts it back to normal text with each line preserved, ready for Sort Lines and Convert to Table. Notes already saved as code blocks: click into them and press Code once.
Two follow-ups to the previous release. (1) Applying bold (or italic, underline, and the other character styles) to colored text flattened the color back to the theme's text color — the display rule fixed last release protected the colored text itself but not styling applied on top of it. Colored text now keeps its color under any combination of character styles. (2) The previous release stopped markdown FILE imports from treating indented lines as program code — but text arriving through AI Chat, OCR, and Describe Image goes through a second, separate markdown converter that still did it, so an indented list of dates from a scanned screenshot still became a code block. That converter now imports indented lines as normal text too, with each line kept on its own line — ready for Sort Lines and Convert to Table. Code blocks written with ``` fences still import as code everywhere. Notes already saved as code blocks keep their stored form; re-running the extraction produces normal text.
Two fixes. (1) Picking a text color or highlight has been writing the color into the note correctly — but a long-standing display rule, added ages ago to keep pasted text readable after theme changes, was flattening every color back to the theme's text color on screen, in every theme. Deliberate colors and highlights are now shown; the readability protection still applies to everything else. (2) Importing markdown treated any line indented by four spaces as program code — an indented list of dates became a code block, where formatting isn't possible. Indented lines now import as normal text; code blocks written with ``` fences still import as code.
Three related dispatch faults. (1) Bold, italic, underline and the other character buttons could silently apply to the wrong note: when editing a note type that uses the classic editing surface (such as checklists), the buttons were routed to the most recently used rich-text note instead — the note being edited never changed. They now reach the note you are actually in. (2) Inside a code block, character formatting is not possible by design (code blocks hold plain program text — the same rule as in other editors); the buttons used to do nothing silently, and now show a brief notice explaining this. (3) The selection toolbar's “More formatting” menu could lose its connection to the editor after the bottom ribbon hid, making Superscript, Subscript and friends dead — they now always reach the editor.
The line tools counted paragraphs and list items — a code block holding many lines, or a paragraph whose lines were separated with Shift+Enter, counted as a single line. So on exactly the notes these tools are most useful for (pasted logs, date lists, delimited data), the selection-toolbar tiles stayed greyed out and the dialogs reported nothing to sort. Both cases are now recognized: lines inside a code block sort in place, Shift+Enter lines sort with their formatting intact, and Convert to Table reads them all. Works from every entry point and with selections inside the block.
The two text tools only recognized notes stored internally as the newer 'text' type. Text notes in existing notebooks — including notes created from markdown import and most notes in real-world documents — are stored as the equivalent legacy 'richtext' type, and on those the tools quietly declined with a brief notice (or the menu entries were missing entirely). Both stored types are now accepted everywhere the tools appear: the formatting ribbon, the note options menu, the selection toolbar, and the Ctrl+K command palette.
Choosing 'Sort lines…' or 'Convert to table…' from a text note's options menu (the ⋮ popup or the note's edit ribbon) did nothing — the dialog was dismissed the instant it opened by the menu's own closing step, and the menu stayed on screen. The menu now closes before the action runs, so both dialogs open normally. The same tools launched from the formatting ribbon's Sort and ToTable buttons or the Ctrl+K command palette were unaffected.
Internal release: the Long Context benchmark now sizes its test prompts using the model's own tokenizer (previously a fixed estimate could overshoot the context window by ~30% on some models, invalidating results), scales its time budgets to the machine's measured speed in its current thermal state, and reports thermal throttling between the start and end of each model's run. No changes to the app itself.
On systems where Gemma 4 models run on the CPU (Intel integrated-graphics machines, per the 5.141.82 policy), the working context is now capped at 8,192 tokens via the model catalog. Measurements showed CPU prompt ingestion runs around 20 tokens per second on this hardware — an uncapped 32K-token request could take close to half an hour, which feels like a hang. The cap keeps every request bounded and predictable; long documents continue to work through the app's chunked AI pipeline, which operates comfortably within the cap. Systems with a dedicated GPU are unaffected and keep the full context. The internal Long Context benchmark was also made more truthful: runs that fail because tasks exceed the time budget now report 'Timeout-Bound' instead of the misleading 'Severe Context Loss', tiers that cannot complete are skipped instead of grinding for hours, and capped tiers are labeled as capped.
Beta 5.141.82 shipped with a coding fault that quietly disabled the AI memory/offload planner on every model start: all models fell back to a legacy full-GPU path, which meant the new Gemma-4-to-CPU routing on Intel integrated graphics never engaged, and Gemma 4 models could still monopolize the GPU — including failing long-context work outright on that path. Fixed: the planner runs again for every model, Gemma 4 models genuinely route to the CPU on Intel iGPU systems (where long-context features also work correctly, since the GPU-specific safety restrictions do not apply on the CPU path), and all other models keep their normal GPU plans. Discrete NVIDIA/AMD/Arc systems remain unaffected.
On laptops and desktops whose only GPU is Intel integrated graphics (Meteor Lake / Arc iGPU class, Windows and Linux), Gemma 4 models no longer offload to the GPU. Benchmarks showed GPU offload of this family on these chips is slower than the CPU (Gemma 4 12B managed 2.1 tokens/s while pinning the GPU near 86% and 90+°C; the same class of model on CPU reached 2.9 tokens/s) and could destabilize under sustained load. A metadata quirk in one 12B model file had let it slip past the existing safeguard — the routing is now independent of model-file metadata, driven by the model catalog with a built-in family fallback. Systems with a dedicated NVIDIA, AMD, or Intel Arc discrete GPU are unaffected, as are all other model families.
The editing improvements from this series are now fully discoverable: 'Convert Table to Text/CSV' joins Sort Lines and Convert to Table in the Ctrl+K command palette; the Help guide gains a Text Tools & Drag-and-Drop section; and the demo notebook's Text Notes & Formatting page adds three try-it notes — a sortable meeting list, a CSV-to-table exercise, and a guide to the single left drag handle. Documentation set (GUI spec amendment, editing scope boundary, SRS and Cross-Reference updates) ships alongside this release.
During a drag, the semi-transparent ghost of the item could be pulled into the page layout by the item's own styling rules instead of floating above it — compressing all three panes and leaving a black band across the lower part of the window until the drop. The ghost's geometry is now pinned so it always floats over the interface. Purely cosmetic: drops, dialogs, and reordering were unaffected.
Internal drag and drop — reordering notes and pages, moving notes to other pages, moving pages and notebooks to other notebooks — no longer uses the browser's native drag system, whose Linux cursor pipeline kept showing a red 'no drop' circle no matter what the app told it. Drags are now driven directly by pointer events: press the drag handle, move a few pixels, and a ghost of the item follows the cursor with a steady grabbing cursor the whole way. The outcome pill, target highlights, and insertion indicators work as before, and dropping calls the exact same actions — reorder, or the move/copy dialog. Invalid spots simply show no highlight and no pill; there is no forbidden cursor anywhere, on any platform. Escape cancels a drag in progress. Dragging files in from your computer, and Android long-press reordering, are unchanged.
Notes previously showed two different drag affordances: the three-line handle on the left of the header and a six-dot handle on the right — one from each of the app's two drag systems. The left handle is now the single drag surface for notes, matching the page drag handle in position and symbol, and it drives everything the right handle used to: reordering with the slot indicator, moving notes to other pages, and Android long-press reordering (now on the left handle, with a full-size touch target). Pages also tighten up: with a mouse, page drags now start from the page's drag handle rather than anywhere on the row — on Android, long-press anywhere on the page row still works.
The 5.141.76 arbiter was being silenced before it could speak: the pane-wide drop zones stop drag events from ever reaching it, so the blocked cursor persisted across the panes. The arbiter now intercepts drags ahead of every other handler, which no handler can block. The move cursor and the outcome pill now hold across the entire journey, and the blocked cursor is set explicitly — not left to chance — over the few genuinely invalid spots. Verified in an automated browser run against the real build. Drop behavior remains untouched.
Dragging a note or page used to show the red 'no drop' cursor across much of the journey even though the drop would work — two internal drag systems each answered only for their own targets. A single drag session now arbitrates every internal drag: the move cursor stays on for the entire valid path, a small pill next to the cursor names the outcome before you commit ('Move / copy — choose on drop' over another page or notebook, 'Move here' for reordering), valid targets highlight consistently, and the blocked cursor now appears only where a drop is genuinely impossible — such as a note over its own page entry. Drop behavior itself is unchanged: reordering, the move/copy dialog, and dragging files in from your computer all work exactly as before.
Text notes gain two long-requested tools. Sort lines orders the lines of a note or a selection — A to Z, Z to A, numeric, natural (2 before 10), reverse, or date-aware, which understands lines beginning with dates like 2026-07-04 09:30 and sorts them chronologically — with options for case-insensitive sorting and removing duplicate or empty lines; while editing, the whole sort is one single Undo step. Convert to table turns selected lines (or a whole note) into a proper table note: the delimiter is detected automatically (commas, tabs, semicolons, pipes, multiple spaces — or date-shaped lines like 2026-07-04 09:30 Meeting notes, which split into date, time, and subject columns), with a live preview, a first-row-is-header toggle, and the choice to insert the table after the note or replace the original (undoable from the toast). Tables gain the reverse: Convert to text/CSV. Find everything on the format row, the selection toolbar’s … menu, the note’s ⋮ menu, and the command palette.
When you select text, the small floating toolbar (bottom bar on Android) has been completed and unified with the main formatting row. It now carries Strikethrough and a Text color picker alongside Bold, Italic, Underline, Highlight, Link, and Checklist, uses the same icons as the formatting row so the two read as one system, and gains a trailing … button opening a small menu with the rarer selection options: superscript, subscript, inline code, a custom highlight color, and clear formatting. Everything works in regular text notes and inside table cells alike, and your selection is preserved through every action — including a round trip through the … menu.
Publishing preparation: every program file inside the Windows package is digitally signed through a cloud signing service, and the build previously requested a fresh security token for every single file — dozens of rapid requests, where one momentary network or service hiccup aborted the whole build after most files had already signed successfully. The build now requests one token per build and reuses it, retries briefly if the service stumbles, and refreshes automatically if a token stops being accepted mid-run. A new self-check also lets the signing credentials be verified at any time without running a build.
Follow-up hardening from the same build-system audit. When the packaging check finds the bundled AI engine isn’t the version it should be — which can happen when the build folder is synced between machines and an older copy briefly reappears — the build now replaces it with the correct engine automatically instead of stopping with instructions. Separately, the step that embeds sign-in credentials could fail to start yet the build continued anyway; any failure there now stops the build outright, at every security level, so a package can never ship with sign-in disabled.
Two build-system checks reported false alarms on every single build: an “incomplete application” warning triggered by an ordinary word in an internal change log, and a credentials check that flagged a harmless internal marker belonging to the text editor’s markdown engine. Both checks now look for exactly what they are meant to find — and the incomplete-application check, which previously only printed a message and continued, now stops the build outright if it ever detects the real thing. Nothing about the shipped application changes; this hardens the process that packages it.
A build-system audit found the Linux package carrying a correctly working but older AI inference engine than its label claimed — the version file had been updated without the engine itself being replaced, and the packaging check trusted the label. The check now asks the engine directly for its version and refuses to package on a mismatch, and the download step verifies each freshly installed engine before certifying it. No user action needed; the affected engine was the same fully functional version shipped in earlier releases.
Two fixes for editing long or structured notes. First, the formatting toolbar attached to the note you are editing used to scroll away with the note — in a note taller than the window you had to scroll back up for every Bold or font change. It now rides at the top of the note and pins to the top of the editor pane as you scroll deeper, staying within reach for the whole edit. Second, clicking into a table cell used to replace the entire formatting toolbar with table controls, making Bold, Italic, fonts, and colors vanish. The character formatting controls now stay put, with the table’s row and column controls appearing alongside them.
Uploading media larger than 2 GB to Google Drive failed with a file-too-large error — files of any size Drive accepts (up to 5 TB) now stream to your account without loading them into memory, for both new uploads and updates to existing files. Production packages also shipped without their product-identity file, silently falling back to a default; the build now stages it correctly and refuses to package without it. Internal version reporting in diagnostics logs is also unified to the real application version.
A delivery check now guarantees that every new version documents itself on this page before it can ship — the newest entry here will always match the version you are running.
This page fell behind again during a rapid run of releases, despite the standing rule that each release documents itself. All thirteen missing releases below are now covered, and the rule is back in force: the newest entry on this page always matches the version you are running.
If you checked “Don’t show on startup,” the Quick Start Guide could become impossible to find again on some platforms. It now has a permanent place in Help & Info (between Help Guide and Load Demo File) and in the Android Help menu. The startup checkbox now means exactly what it says: it only stops the automatic opening at launch.
On phone screens the Quick Start Guide’s sections could collapse into overlapping slivers with their text spilling behind neighboring cards. Each section now renders at its full height and the guide scrolls normally. Desktop was unaffected.
Three fixes from one field session. The “listening” indicator used to appear about half a second before the microphone was actually live, silently losing the start of what you said — the microphone now starts early so it is truly live the moment the indicator appears. Speaking with background noise suppression on your device could make natural mid-sentence pauses count as the end of your turn — the voice detector now rides through those dips. And the per-device voice calibration no longer drifts downward from quiet recordings.
On desktop, the hands-free mode could re-open the microphone while the assistant was still speaking, hear its own voice through the speakers, and reply to it. The microphone now stays closed until every part of the spoken response has actually finished playing. Note: talking over the assistant is not picked up either — the conversation is strictly turn-based.
A build-system correction with no changes to the app itself.
A build-system correction with no changes to the app itself.
The speech engine previously reloaded its voice model for every sentence. It now stays loaded for the whole session, so the first spoken words of a response arrive sooner and short sentences no longer leave small gaps between them.
Conversational Mode on Android previously needed a tap on Done (or a 12-second wait) to finish each turn. Android now detects the end of your sentence from the audio itself, just like desktop. If a turn is finished but nothing was understood, the assistant now says so instead of silently waiting again, and the screen stays responsive while your speech is transcribed.
A recording that contained no actual words (background noise, an accidental press) could gradually teach the voice detector that your speaking voice is as quiet as the room — degrading hands-free turn detection over time. Only recordings with real words now contribute to that per-device learning.
On phones with mixed fast/efficiency processor cores, transcription was spreading its work across all cores and being held back by the slow ones. It now uses only the fast cores — the same fix the AI models received earlier — making each spoken turn noticeably quicker to process.
Two overlapping requests to prepare the AI model could each load a full copy into memory, and on phones the doubled memory use could get the app terminated by the system — sometimes appearing afterwards as a microphone that had gone silent. Only one copy can now ever load; the duplicate request simply waits and reuses it.
AI Chat gains a hands-free mode: tap the conversation button next to the microphone and just talk. The assistant listens, answers aloud, and listens again — no typing or button presses between turns. Say “exit conversation mode” (or start typing) to leave, and adjust the pause length and spoken cues under Voice & Speech settings. A persistent indicator always shows when the microphone is live.
Desktop dictation gained real-time speech/silence detection and rolling interim transcripts under the hood — the foundation for Conversational Mode’s natural turn-taking in the next release. Push-to-talk behavior was unchanged.
The line at the bottom of a saved AI Chat note — which model answered and how fast — previously showed speed figures only for some kinds of responses. Every saved response now reports the same details in the same format, and always names the exact model that produced that specific answer, even if you switched models afterwards.
Creating a note from an earlier AI Chat answer after you had already asked a follow-up question could save the note without your attached image, and could quote the wrong question above the answer. Each answer’s Create Note button now remembers exactly which question — and which attachments — produced it. And when you make a note from a follow-up answer about an image, that image is now included too, marked as coming from earlier in the conversation.
This page had fallen a few releases behind the app. It now covers every release, and from this version forward each release documents itself here — including the one you are running right now.
On shorter windows and phone screens, some dialogs — the PDF transcription and description options among them — could extend past the edges of the screen, leaving the buttons at the bottom unreachable. Every dialog in the app now fits within the screen and scrolls internally whenever its content is taller than the window.
In our production builds, text written by the AI — document summaries in particular — could appear as one long unformatted block, with stray “#” and “*” characters where headings, bold text, and lists should have been. A step in our release build process was interfering with the editor’s formatting engine; that step now leaves it untouched, so release builds format AI output exactly as intended. The app is also several megabytes smaller as a result.
The Errata page introduced in 5.141.46 wasn’t reachable from the Help & Info menu on the current interface — the button had only been added to older menu locations. It now appears everywhere Help does.
The Help menu now includes an Errata page listing what changed in each release, newest first — the same information as this changelog, right inside the app. It replaces the old Version History page, which had fallen far out of date.
The speech-recognition engine now stays ready between dictations instead of starting from scratch every time, so after your first dictation of a session, the next ones begin transcribing noticeably sooner. If the engine is ever unavailable, dictation quietly uses the original method instead — nothing changes in how you work.
While you dictate, NotesXML now receives your recording in finer increments. On systems where audio previously arrived in coarse bursts, in-progress transcription updates could lag behind; they now track your speech in near real time.
When reading AI responses aloud as they arrived, the voice could occasionally say “asterisk” out loud where a bold or italic phrase happened to be split across sentences. Formatting characters are now cleaned up reliably — and as part of the same fix, technical names written with underscores read naturally, tables read as brief pauses instead of “pipe,” and simple arithmetic like 3 * 4 is spoken as “3 times 4.”
Hovering over an item in the AI menu — opened from a note, page, or notebook — could make its text unreadable in several of our light color themes. This also affected the Translate and Convert Code language pickers reached from that menu. Both are now fixed across every theme.
Fixed a rare issue where opening the same notebook a second time — for instance, clicking it twice while distracted — could cause your most recent changes to be silently lost when you saved. NotesXML now recognizes when a notebook is already open and switches you to it instead of opening a duplicate copy. As an extra safeguard, if a notebook ever changes on disk after you've opened it, NotesXML will ask before saving over those changes rather than overwriting them automatically.
When you turn an AI Chat conversation into a note, any images or files you attached are now saved as part of the note alongside the AI's response, instead of being left out.
A “#” character appearing in the middle of AI-generated text — from OCR, image descriptions, or AI Chat — was sometimes mistaken for heading formatting and rendered as a large heading. It now correctly stays part of the sentence.
Settled on the curated set of on-device AI models we're shipping at launch, and corrected an internal image-processing setting that was quietly reducing the accuracy of vision-based AI actions like Describe Image.
On Intel Arc integrated graphics, some AI models — including the Gemma family — don't yet run reliably on the GPU, so NotesXML automatically routes them to the CPU instead. Along the way we fixed a crash that could recur every time you tried again, and a separate issue where "CPU-only" mode wasn't fully disabling GPU use, which was quietly throttling CPU speed. Other models, like our default Ministral 3, already run well on Intel Arc's GPU.
Summarizing an unusually long note or document could overflow the AI's working memory and crash the app. Content is now measured and split more carefully before being sent to the AI, so long documents summarize reliably.
AI actions on Link and Video Link notes now work from a more consistent set of source content and return more focused results.
On Android, a model download could fail without any error message, leaving the model unusable with no explanation. Failed downloads are now detected and reported properly.
Math equations written with our math editor were sometimes missing from printed pages and PDF exports, even though they displayed correctly on screen. They now print and export exactly as shown in the editor.
The Microsoft Store edition of NotesXML now supports purchasing and activating a Professional license directly through the Store, in addition to entering a license key.
In rare cases, other elements on the setup screen could interfere with PIN entry during the security setup step. The PIN field is now fully isolated from anything else on the page.
NotesXML for Android no longer requests photo or video media permissions. Attaching images and files now goes entirely through Android's built-in file picker, so the app can do its job with less access to your device.
This page currently covers every release from 5.141.14 through 5.151.6. Have a question about a specific fix? Contact support.