NotesXML AI Model Benchmark Results & Recommendations
IWV Digital Solutions LLC | measured against catalog v2026.08.13.02 (10 models) | 13 August 2026
These results measure catalog v2026.08.13.02 (10 models), benchmarked 13 August 2026. Every figure below is from that run.
The shipping catalog is now v2026.09.06.02 with 16 models. Gemma 4 E2B, Mellum2 12B-A2.5B, Devstral Small 2 24B, Muse Glimmer 30B, Laguna XS 2.1 33B-A3B and Nemotron 3.5 Lightning 30B-A3B were added after this run and are not in the tables below. A full re-benchmark is pending rather than an extension of these tables, for two reasons: the reference workstation now runs two GPUs, so throughput figures are not comparable with the single-GPU numbers here; and several evaluation suites have been rebuilt since August — tool calling now uses the model's native tool-call API with a multi-turn agentic loop instead of JSON-in-prompt, the prompt sweep covers 111 prompts, and OCR is accuracy-scored against fixture ground truth. Mixing the two would produce a table that looks comparable and is not. See Products → AI Models for the current catalog.
Overview
At the time of this run, NotesXML's Professional tier shipped with 10 local AI models (it now ships 16). They span every hardware tier, from phones and low-RAM devices to enthusiast workstations with discrete GPUs. This article gives the full benchmark results for all 10 catalog models, the Harmonic Mean Score (HMS) ranking, and the efficiency-frontier analysis.
Catalog v2026.08.13.02 (August 2026): The catalog is built by running the full benchmark suite against a broad set of candidate models, computing the Pareto frontier on TPS vs average quality, and applying a utility floor to drop models below viable capability. The ten shipping models are LFM2.5 VL 1.6B, Ministral 3 3B, LFM2.5 VL 3B, Gemma 4 E4B, Ministral 3 8B, Gemma 4 12B, Ministral 3 14B, GPT-OSS 20B, Gemma 4 26B-A4B and Gemma 4 31B. Ministral 3 14B is retained for Intel iGPU/GPU systems, where it runs on the Intel GPU while the Gemma 4 models run CPU-only. Nine of the ten are vision-capable — every model except GPT-OSS 20B. Free tier: LFM2.5 VL 1.6B (recommended default on both platforms) and Ministral 3 3B, both vision-capable. Llama 3.2 3B Instruct and Phi-4 Mini 3.8B were retired in the 12 August revision; LFM2.5 VL 3B and GPT-OSS 20B hold their places on the frontier, LFM2.5 VL 3B having replaced LFM2 VL 3B on 13 August.
Refreshed August 13, 2026 — catalog v2026.08.13.02. LFM2.5 VL 3B (Liquid AI, vision) replaces LFM2 VL 3B in the Lightweight tier after a head-to-head re-run: it is faster (279.3 vs 270.3 TPS) and marginally higher in average quality (90.3% vs 90.1%), trading some coding strength for stronger exam reasoning. The 12 August revision had already added GPT-OSS 20B and retired Llama 3.2 3B Instruct and Phi-4 Mini 3.8B. All ten shipping models sit on the frontier, and nine of ten are vision-capable.
| Shipping model | Vision | RAM (GB) | TPS | Quality | HMS |
|---|---|---|---|---|---|
| LFM2.5 VL 1.6B | Yes | 1.3 | 584.8 | 87% | 0.93 |
| Ministral 3 3B | Yes | 3 | 225.1 | 94% | 0.97 |
| LFM2.5 VL 3B | Yes | 2.9 | 279.3 | 90% | 0.95 |
| Gemma 4 E4B | Yes | 6 | 118.8 | 97% | 0.99 |
| Ministral 3 8B | Yes | 7 | 116.0 | 97% | 0.99 |
| Gemma 4 12B | Yes | 8 | 70.7 | 99% | 0.99 |
| Ministral 3 14B | Yes | 11 | 74.6 | 98% | 0.99 |
| GPT-OSS 20B (MoE) | — | 18 | 141.5 | 95% | 0.97 |
| Gemma 4 26B-A4B | Yes | 22 | 51.4 | 100% | 1.00 |
| Gemma 4 31B | Yes | 26 | 8.5 | 100% | 0.29 |
For reference, the models retired in this revision score: Llama 3.2 3B Instruct — HMS 0.89 (80% quality, 257.0 TPS, text-only); Phi-4 Mini 3.8B — HMS 0.93 (86% quality, 229.0 TPS, text-only). Both remain usable if already downloaded.
All AI inference in NotesXML runs entirely on your device. No internet connection is required. No data is transmitted.
Benchmark Deep-Dives
Each benchmark in the suite has its own article. The article covers how the benchmark works, the full question or task list, how answers are scored, and the research behind its design. This page gives the scores; the deep-dive articles explain what the scores mean.
NotesXML Benchmark Suite
NotesXML uses eight purpose-designed benchmarks to evaluate models as they actually run on local hardware — testing the combination of model capability and local inference performance that users experience in practice.
Long Context Benchmark (100 points)
Evaluates a model's ability to process, retain, and reason over extended input sequences. Passages with embedded facts at varying context depths are presented; the model is queried on details requiring genuine retention rather than positional pattern-matching. Critical for note-taking applications where users ask AI to summarize or analyze multi-page documents.
| Score Range | Level |
|---|---|
| 80–100 (80–100%) | Strong Retention |
| 60–79 (60–79%) | Functional Retention |
| 40–59 (40–59%) | Partial Retention |
| Below 40 (< 40%) | Limited Retention |
Tool Calling Benchmark (100 points)
Evaluates a model's ability to correctly invoke structured function calls. Tests cover simple single-tool invocations, multi-step chains, parallel calls, and error recovery — all capabilities underpinning NotesXML's AI action system.
| Score Range | Level |
|---|---|
| 90–100 (90–100%) | Expert Tool Use |
| 80–89 (80–89%) | Advanced Tool Use |
| 70–79 (70–79%) | Reliable Tool Use |
| Below 70 (< 70%) | Basic Tool Use |
Model Benchmark — In-App Task Suite (Grade A–F + Harmonic)
Runs the seven tasks NotesXML performs in practice and grades each A–F: six text tasks — summarization, auto-titling, grammar and spelling correction, English→Spanish translation, action-item extraction, and Markdown/XML generation — plus a vision OCR task for vision-capable models. The per-task grades are combined into a normalized quality score (exported as the AvgScore field) and an overall letter grade; the Harmonic score pairs that quality with inference speed, rewarding models that are both correct and fast. Chat-template handling, stop-token behavior, and output-format integrity are scored within each task.
HEE — Human Education Evaluation (50 points)
Evaluates broad factual knowledge across mathematics, science, history, literature, geography, and reasoning — drawn from secondary school through postgraduate level, weighted toward the upper end. The closest NotesXML benchmark to the academic MMLU family.
| Score Range | Level |
|---|---|
| 45–50 (90–100%) | Post-Graduate / Mastery |
| 40–44 (80–89%) | Undergraduate Level |
| 30–39 (60–79%) | High School Graduate |
| Below 30 (< 60%) | Below Standard |
PLE — Professional License Exam (200 points)
200 questions across ten professional fields: Law (MBE), Medicine (USMLE), Engineering (PE), Psychology (EPPP), Nursing (NCLEX-RN), Real Estate, Finance (CPA), Architecture (ARE), and Cybersecurity (CISSP). Tests not only factual recall but multi-step logical deduction and contextual synthesis under professional-exam constraints.
| Score Range | Level |
|---|---|
| 190–200 (95–100%) | Expert / Mastery |
| 170–189 (85–94%) | Expert / Mastery |
| 160–169 (80–84%) | Proficient |
| 140–159 (70–79%) | Developing |
| Below 140 (< 70%) | Below Competency |
PhD Philosophy — PhD-Level Logic & Philosophy Comprehensive Exam (100 points)
100 questions spanning metalogic, modal logic, philosophy of language and mind, epistemology, metaphysics, and continental phenomenology. Designed so that correctly answering requires understanding how theories mechanically function, not merely pattern-matching to familiar names. The most intellectually demanding single benchmark in the suite.
| Score Range | Level |
|---|---|
| 90–100 (90–100%) | PhD Mastery / Expert |
| 80–89 (80–89%) | Advanced |
| 70–79 (70–79%) | Proficient |
| Below 70 (< 70%) | Below PhD Level |
SDB-100 — Synthetic Deduction Benchmark (100 points)
10 questions testing multi-step deductive reasoning using synthetic ontologies that cannot exist in any training corpus — novel self-contained universes governed by arbitrary but logically absolute rules. By stripping away semantic priors, the benchmark forces genuine structural reasoning rather than pattern completion. The most discriminating and contamination-resistant benchmark in the suite.
| Score Range | Level |
|---|---|
| 8–10 (80–100%) | Advanced Deductive Reasoning |
| 6–7 (60–79%) | Proficient |
| 4–5 (40–59%) | Developing |
| Below 4 (< 40%) | Below Standard |
Full Results
NotesXML Benchmark Results
Full benchmark sweep for all 10 catalog models — master catalog refreshed 13 August 2026 with the updated tool-calling suite, on Ryzen 7 7800X3D (Linux). Rows are sorted by HMS (the speed/quality score). ★ marks the models on the Pareto efficiency frontier of the 76-model sweep — all 10 catalog models qualify in this revision. A dash means the benchmark does not apply to that model — GPT-OSS 20B is text-only and Gemma 4 31B was not run on the vision suite.
| Model | Tier | Avg TPS | Grade | HEE /50 | PLE /200 | PhD /100 | SDB-100 /100 | Code /100 | LongCtx /100 | Tool /100 | Avg Score | HMS |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B | Desktop Pro | 51.4 | A | 100% | 99% | 100% | 80% | 99% | 85% | 94% | 100% | 1.00 |
| Gemma 4 E4B | Enhanced | 118.8 | A | 96% | 92% | 99% | 46% | 86% | 77% | 97% | 97% | 0.99 |
| Ministral 3 8B | Enhanced | 116.0 | A | 96% | 95% | 96% | 72% | 95% | 86% | 83% | 97% | 0.99 |
| Gemma 4 12B | Enhanced | 70.7 | A | 98% | 96% | 99% | 54% | 97% | 81% | 93% | 99% | 0.99 |
| Ministral 3 14B | Advanced | 74.6 | A | 98% | 95% | 96% | 47% | 92% | 87% | 78% | 98% | 0.99 |
| GPT-OSS 20B | Desktop Pro | 141.5 | A | 100% | 93% | 96% | 69% | 93% | 82% | 4% | 95% | 0.97 |
| Ministral 3 3B | Lightweight | 225.1 | A | 88% | 89% | 94% | 41% | 88% | 88% | 84% | 94% | 0.97 |
| LFM2.5 VL 3B | Lightweight | 279.3 | A | 88% | 83% | 89% | 41% | 63% | 30% | 0% | 90% | 0.95 |
| LFM2.5 VL 1.6B | Ultra-Light | 584.8 | A | 78% | 74% | 90% | 28% | 67% | 79% | 4% | 87% | 0.93 |
| Gemma 4 31B | Desktop Pro | 8.5 | A | 100% | 100% | 100% | 51% | 61% | 92% | 91% | 100% | 0.29 |
Avg TPS = tokens per second (text tasks only). Grade = Model Benchmark overall grade (all 10 models in this run Grade A). Avg Score = qualityMean, the unweighted mean of the normalized benchmark scores that apply to the model (Model Benchmark text and vision, HEE, PLE, PhD, SDB-100, Code, LongCtx, Tool). HMS = harmonic mean of min(TPS/50,1) and Avg Score.
Performance vs Token Interval — The Efficiency Frontier
The plot shows each model’s quality against its token interval (1/TPS, milliseconds per generated token — lower is faster). The dashed line is the Pareto efficiency frontier: the models that no other model beats on both speed and quality at once. All 10 catalog models sit on the frontier in the v2026.08.12.01 revision. Gemma 4 31B plots far left at 8.5 TPS — a VRAM artifact of the 16 GB reference rig, not a property of the model: see the note under the chart.
| Model | Avg TPS | TPS per GB RAM | Avg Score | Frontier |
|---|---|---|---|---|
| LFM2.5 VL 1.6B | 584.8 | 449.85 | 87% | ★ YES |
| LFM2.5 VL 3B | 279.3 | 96.31 | 90% | ★ YES |
| Ministral 3 3B | 225.1 | 75.03 | 94% | ★ YES |
| GPT-OSS 20B | 141.5 | 7.07 | 95% | ★ YES |
| Gemma 4 E4B | 118.8 | 19.80 | 97% | ★ YES |
| Ministral 3 8B | 116.0 | 16.57 | 97% | ★ YES |
| Ministral 3 14B | 74.6 | 6.78 | 98% | ★ YES |
| Gemma 4 12B | 70.7 | 8.84 | 99% | ★ YES |
| Gemma 4 26B-A4B | 51.4 | 2.34 | 100% | ★ YES |
| Gemma 4 31B | 8.5 | 0.33 | 100% | ★ YES |
Sorted by TPS descending (fastest first). All 10 models in this run are on the strict Pareto efficiency frontier of the 76-model refreshed run.
Harmonic Mean Score — Single-Number Speed/Quality Ranking
The Harmonic Mean Score (HMS) folds speed and quality into one number. Higher is better. The harmonic mean punishes imbalance: a model that is fast but weak, or strong but slow, scores lower than one that is good at both. Speed is capped at 50 TPS. Past that point a model already streams faster than most people read in a chat window, so extra speed adds nothing to the experience.
Methodology
- Speed component:
speed_norm = min(TPS / 50, 1.0) - Quality component:
quality = qualityMean— the unweighted mean of all eight normalized benchmark scores. Code Benchmark is now included. - Harmonic mean:
HMS = 2 × speed_norm × quality / (speed_norm + quality)
HMS Ranking (all 10 catalog models)
| Rank | Model | Tier | Avg TPS | Avg % | speed_norm | quality | HMS |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 26B-A4B | Desktop Pro | 51.4 | 100% | 1.000 | 1.00 | 1.00 |
| 2 | Gemma 4 E4B | Enhanced | 118.8 | 97% | 1.000 | 0.97 | 0.99 |
| 3 | Ministral 3 8B | Enhanced | 116.0 | 97% | 1.000 | 0.97 | 0.99 |
| 4 | Gemma 4 12B | Enhanced | 70.7 | 99% | 1.000 | 0.99 | 0.99 |
| 5 | Ministral 3 14B | Advanced | 74.6 | 98% | 1.000 | 0.98 | 0.99 |
| 6 | GPT-OSS 20B | Desktop Pro | 141.5 | 95% | 1.000 | 0.95 | 0.97 |
| 7 | Ministral 3 3B | Lightweight | 225.1 | 94% | 1.000 | 0.94 | 0.97 |
| 8 | LFM2.5 VL 3B | Lightweight | 279.3 | 90% | 1.000 | 0.90 | 0.95 |
| 9 | LFM2.5 VL 1.6B | Ultra-Light | 584.8 | 87% | 1.000 | 0.87 | 0.93 |
| 10 | Gemma 4 31B | Desktop Pro | 8.5 | 100% | 0.170 | 1.00 | 0.29 |
Reading the HMS Ranking
On the August 2026 run, Gemma 4 26B-A4B leads at HMS 1.00. With the speed cap at 50 TPS its 51.4 TPS just clears the cap, so its 100% average quality is no longer discounted — it tops the ranking rather than sitting mid-pack. Four models tie immediately behind it at 0.99: Gemma 4 E4B, Ministral 3 8B, Gemma 4 12B and Ministral 3 14B. GPT-OSS 20B and Ministral 3 3B follow at 0.97.
With the cap at 50 TPS, every model except Gemma 4 31B reaches speed_norm = 1.0. For those nine models, HMS is driven almost entirely by quality. This is by design: once a model streams faster than you can read, quality is what separates them. The fastest model in the catalog, LFM2.5 VL 1.6B (584.8 TPS), ranks ninth of ten at 0.93 because its 87% quality is the lowest of the shipping models — speed alone cannot lift it past larger models that also clear the cap.
The speed cap decides the bottom of the table, not the top — and on this rig the cap is VRAM, not the model. Gemma 4 31B is the clearest case: it carries the highest raw quality in the sweep, but 8.5 TPS gives it a speed_norm of just 0.17 and an HMS of 0.29. That figure is a property of the reference rig, not of the model. The reference GPU holds 16 GB of VRAM and Gemma 4 31B needs about 26 GB at peak, so the weights cannot fit and the remainder spills to system RAM, where every token pays host-memory bandwidth. Dense models are hit hardest by that spill because every parameter is touched on every token, whereas the Mixture-of-Experts models above it in the table activate only a fraction per token and degrade far more gently. On a GPU with enough VRAM to hold the model, Gemma 4 31B is expected to perform substantially better, and its HMS on this table should be read as "on 16 GB VRAM" rather than as a ceiling. It carries the highest raw quality score in the entire sweep. LFM2.5 VL 1.6B sits at the opposite corner: the smallest model of the shipping ten at 87% quality, but so far above the cap (584.8 TPS) that HMS 0.93 places it ninth of ten only because everything above it is larger.
Key Observations
All 10 catalog models sit on the Pareto frontier of the 76-model August 13 sweep. The v2026.08.13.02 revision retired the two dominated entries from the previous catalog — Llama 3.2 3B (80% at 257.0 TPS) and Phi-4 Mini 3.8B (86% at 229.0 TPS), both text-only and both outscored in their own speed class by Ministral 3 3B and LFM2.5 VL 3B, which are vision-capable — and added the two frontier models that replaced them: LFM2.5 VL 3B (90% at 279.3 TPS, the fastest model above the Ultra-Light tier) and GPT-OSS 20B (95% at 141.5 TPS, the fastest large model in the catalog).
Mid-size Enhanced models fill the top of the HMS ranking. Gemma 4 E4B, Ministral 3 8B and Gemma 4 12B (HMS 0.99) clear the 50-TPS cap and deliver 97–99% quality — the best speed/quality balance in the catalog, behind only the desktop-only 26B MoE. Among models fast enough to stream comfortably, the ranking rewards quality.
Gemma 4 26B-A4B (MoE) leads on quality among models that clear the speed cap. It takes the top spot on almost every knowledge and reasoning benchmark: 100% average quality, 100% HEE, 99% PLE, 100% PhD Philosophy, 80% SDB-100 (catalog leader), and 99% Code. With vision and a 256K-token context, it is the highest-quality model you can run at a readable speed on the reference rig — only Gemma 4 31B scores higher overall, and it does so at 8.5 TPS. Desktop Pro only.
Vision now spans the catalog — including the free tier. Nine of the ten models in this run are vision-capable: LFM2.5 VL 1.6B, Ministral 3 3B, LFM2.5 VL 3B, Gemma 4 E4B, Ministral 3 8B and Gemma 4 12B run cross-platform; Ministral 3 14B, Gemma 4 26B-A4B and Gemma 4 31B are desktop-only. GPT-OSS 20B is the only text-only model, chosen for throughput on long-form writing and reasoning. Both free-tier models read images and handwriting. Long Context retention is led by Gemma 4 31B (92%) and Ministral 3 3B (88%); Tool Calling by Gemma 4 E4B (97%) — note that LFM2.5 VL 1.6B and GPT-OSS 20B both score 4% on Tool Calling, and LFM2.5 VL 3B scores 0%, so none of the three is a tool-driving choice.
Model Recommendations by Hardware Tier
Ultra-Light — Up to ~1.5 GB RAM Available for AI
Free-tier default: LFM2.5 VL 1.6B (HMS 0.93) — the recommended default on both desktop and Android and the fastest model in the shipping catalog at 584.8 TPS, in 1.3 GB of RAM. Unusually for its size it is vision-capable, so image and handwriting recognition are available on the free tier. It replaces Llama 3.2 1B Instruct (HMS 0.81 on this run), beating it on quality (87% vs 68%) and speed (584.8 vs 559.0 TPS) while adding vision in less RAM. Best for instant note titling, short summaries, quick polish, and on-device image and handwriting reading. Offered under the LFM Open License v1.0 — personal use, and commercial use by organisations under US $10M annual revenue; Ministral 3 3B is the unrestricted free-tier alternative.
Lightweight — ~1.5–4.5 GB RAM Available for AI
Best all-round Lightweight model: Ministral 3 3B (HMS 0.97) — Apache 2.0, 225.1 TPS, vision-capable, and the Electron default. The first vision model in the Free tier. 94% average quality (HEE 88%, PLE 89%, PhD 94%, Code 88%, Long Context 88%, Tool 84%), 256K native context. The best all-round default for note-taking with image support on Lightweight hardware.
Fast text specialist: GPT-OSS 20B (HMS 0.97) — Apache 2.0, 141.5 TPS on the reference rig, text-only. HEE 100%, PhD 96%, Code 93%. The pick for long-form writing, summarization and reasoning when no image input is needed. It runs with reasoning effort set to low, which is what keeps it fast; tool calling is its clear weak spot (4%), so pair it with another model for tool-driven work. Desktop only, ~18 GB peak.
Fastest vision model above Ultra-Light: LFM2.5 VL 3B (HMS 0.95) — 279.3 TPS, 90% quality, vision-capable, 2.9 GB RAM, cross-platform. The speed pick for image-heavy note work on modest hardware. LFM Open License v1.0 (commercial use under US $10M annual revenue).
Ultra-Light vision option: LFM2.5 VL 1.6B (HMS 0.93) — LFM Open License v1.0, 584.8 TPS, vision-capable in 1.3 GB, the recommended default on both platforms and a free-tier model. 87% average quality (HEE 78%, PLE 74%, PhD 90%, Code 67%, Long Context 79%), with Tool Calling at 4% — it reads and writes well for its size but is not a tool-driving model. It replaces Gemma 4 E2B as the small vision option, at less than a third of the RAM.
Enhanced — ~5–9 GB RAM Available for AI
Best speed/quality balance: Ministral 3 8B (HMS 0.99) — Apache 2.0, 116.0 TPS, vision-capable. 97% average quality (HEE 96%, PLE 95%, PhD 96%, Code 95%, Long Context 86%), and the catalog’s runner-up on SDB-100 at 72%. 256K native context. The top speed/quality balance for AI Chat, structured actions, and image analysis on Enhanced hardware.
Tool-calling flagship: Gemma 4 E4B (HMS 0.99) — Apache 2.0, 118.8 TPS, Tool 97% (catalog leader), HEE 96%, PLE 92%, PhD 99%. Vision-capable, 128K native context. The strongest cross-platform choice for tool-driven AI actions.
Highest-quality cross-platform model: Gemma 4 12B (HMS 0.99) — Apache 2.0, 70.7 TPS, 99% average quality (the catalog’s highest outside the two desktop-only Gemma 4 models), Code 97%, PhD 99%, Tool 93%. Vision-capable, 12B dense, 256K native context. The top-quality model that still runs cross-platform, one place off the overall HMS lead.
Intel iGPU / GPU pick: Ministral 3 14B (HMS 0.99, Advanced tier, desktop only) — Apache 2.0, vision-capable, 14B dense, 256K native context, ~11 GB peak RAM. This is the large-model option for Intel systems, running on the Intel iGPU or GPU through Vulkan or SYCL like every other catalog model. On the Ryzen reference rig it measured 74.6 TPS at 98% average quality (HEE 98%, PLE 95%, PhD 96%, Code 92%, Long Context 87%), so on that machine Ministral 3 8B just edges it on the speed/quality frontier. On an Intel GPU, though, the 14B is how you run a large, high-quality vision model at usable speed. Desktop only (Windows and Linux).
Desktop Pro — ~22 GB RAM or 16+ GB GPU VRAM
Highest overall quality: Gemma 4 26B-A4B (MoE) (HMS 1.00) — Apache 2.0, desktop-only, 51.4 TPS. The catalog’s HMS leader: 100% average quality, HEE 100%, PLE 99%, PhD 100%, SDB-100 80% (catalog leader), Code 99%, Tool 94%. 26B total / 4B active Mixture-of-Experts, vision-capable, 256K native context. The highest-quality choice for research, synthesis, and agentic workflows where speed is secondary.
Summary Recommendation Table
| Your Hardware / Use Case | Recommended Model | HMS | Primary Strength |
|---|---|---|---|
| Phone / very low-RAM device (<1.5 GB) — vision in Free tier | LFM2.5 VL 1.6B | 0.93 | Recommended default, both platforms; Free tier; 584.8 TPS in 1.3 GB; vision-capable; LFM Open License v1.0 (commercial use under $10M revenue) |
| Light device, best all-round default — vision in Free tier | Ministral 3 3B | 0.92 | Electron default; vision-capable; Free-tier vision; 256K context; Apache 2.0 |
| Light device, fast vision | LFM2.5 VL 3B | 0.95 | 279.3 TPS; vision; 2.9 GB; cross-platform; LFM Open License v1.0 |
| Desktop, fast long-form text | GPT-OSS 20B | 0.97 | 141.5 TPS; HEE 100%; Apache 2.0; text-only; desktop only; weak tool calling |
| Enhanced desktop — catalog HMS co-leader | Ministral 3 8B | 0.96 | Long Context leader (88%); vision-capable; 256K context; Apache 2.0 |
| Enhanced desktop — tool calling (Pro) | Gemma 4 E4B | 0.94 | Tool 94% (co-leader); vision-capable; cross-platform; Apache 2.0 |
| Enhanced desktop — highest quality cross-platform (vision), HMS co-leader | Gemma 4 12B | 0.95 | Quality 90%; Code 97%; PhD 99%; vision-capable; Apache 2.0 |
| Intel iGPU / GPU — large vision model on Intel hardware (desktop) | Ministral 3 14B | 0.93 | Strongest large vision model measured on Intel hardware; 14B; vision-capable; 256K context; desktop only; Apache 2.0 |
| Desktop Pro — highest quality + vision | Gemma 4 26B-A4B (MoE) | 0.87 | Quality leader (95%); SDB-100 80%; vision-capable; 256K context; Apache 2.0 |
Complete Benchmark Data
Full benchmark results for every model evaluated in the candidate sweep — including models not selected for the catalog — are available in a dedicated reference article.
Notes on Hardware Acceleration
The benchmarks above were collected on a Ryzen 7 7800X3D system running Linux (AI Benchmark Orchestrator v2.3, Beta 5.141.38). NotesXML supports GPU acceleration on NVIDIA CUDA/Vulkan, AMD Vulkan (RADV), and Intel iGPU (Vulkan / SYCL). CPU-only inference remains fully supported on every platform.
- NVIDIA CUDA / Vulkan: Full support. RTX 4060 class and above can run all 10 catalog models. The Desktop Pro tier (GPT-OSS 20B, Gemma 4 26B-A4B, Gemma 4 31B) is best on 16+ GB VRAM at Q4_K_M; Gemma 4 31B needs roughly 26 GB in VRAM to run at full speed.
- AMD Vulkan (RADV): Full support. APU systems (Ryzen 7000 with Radeon 780M) successfully run catalog models through the Lightweight and Enhanced tiers.
- Intel iGPU: Vulkan is supported on Linux and Windows; SYCL is supported on platforms where the Intel oneAPI runtime is available. Every catalog model is GPU-first on Intel hardware; CPU-only inference remains fully supported as a setting you choose (AI Settings → Advanced → Force CPU Only), never as an automatic fallback. Ministral 3 14B is the recommended large model for Intel iGPU/GPU systems on measured speed and quality.
About NotesXML AI
All AI runs locally on your device using llama.cpp. No subscription is needed beyond the one-time Professional license. Models download once and are stored on your machine. No internet is needed when the AI runs. Your notes, questions, and AI responses stay private.
© 2026 IWV Digital Solutions LLC. All rights reserved.
© 2026 IWV Digital Solutions LLC. All rights reserved.
← Back to Articles