10 general datasheets · 102 special-purpose models · rev. 2026.10
Model datasheets
Every open-weight model we track, with its architecture checked against the model's own config.json and its GGUF file sizes read from the Hugging Face API.
Catalogue coverage is limited to the models listed below. Gemma 3 4B and 12B from the setup guide, and the Gemma 4 and Qwen 3.8 families discussed in the cover story, are not supported by our fit calculator yet. Editorial coverage does not supply verified model sizes or fit data. Try the tested CPU setup.
All models
Smallest first. The 24 GB column is a used RTX 3090 at the default quant and 8k context. Speeds: sample or estimate
Model | Parameters | Context | Default quant | File | 24 GB card, 8k | Decode, tok/s | Licence |
|---|---|---|---|---|---|---|---|
| Llama 3.1 8B Instruct Meta | 8.0 B | 128k | Q4_K_M | 4.9 GBVerified | Fits | 93.7Estimate | Llama 3.1 Community License |
| Qwen3 8B Alibaba Qwen | 8.2 B | 32k | Q4_K_M | 5.0 GBVerified | Fits | 90.0Estimate | Apache-2.0 |
| Qwen3 14B Alibaba Qwen | 14.8 B | 32k | Q4_K_M | 9.0 GBVerified | Fits | 54.3Estimate | Apache-2.0 |
| gpt-oss-20b OpenAI | 20.9 B 3.6 B active | 128k | MXFP4 | 12.1 GBVerified | Fits | 245.3Estimate | Apache-2.0 |
| Mistral Small 3.2 24B Mistral AI | 24.0 B | 128k | Q4_K_M | 14.3 GBVerified | Fits | 35.8Estimate | Apache-2.0 |
| Gemma 3 27B IT Google DeepMind | 27.4 B | 128k | Q4_K_M | 16.6 GBVerified | Fits | 31.8Estimate | Gemma Terms of Use |
| Qwen3 30B-A3B Alibaba Qwen | 30.5 B 3.3 B active | 32k | Q4_K_M | 18.6 GBVerified | Fits | 199.8Estimate | Apache-2.0 |
| Qwen3 32B Alibaba Qwen | 32.8 B | 32k | Q4_K_M | 19.8 GBVerified | Tight | 25.6Estimate | Apache-2.0 |
| Llama 3.3 70B Instruct Meta | 70.6 B | 128k | Q4_K_M | 42.5 GBVerified | Needs offload | 1.4Estimate | Llama 3.3 Community License |
| gpt-oss-120b OpenAI | 116.8 B 5.1 B active | 128k | MXFP4 | 63.4 GBVerified | Needs offload | 17.3Estimate | Apache-2.0 |
Special-purpose models
102 open-weight models for one job each: maths, code, vision, documents, speech, images, search, small fast decisions, medicine, tools, languages, forecasting and robots. Each one was checked against the Hugging Face API on 2026-10-13 (parameters and licence). Scores come from one public benchmark per category, cited per row; where nobody has published a comparable number we say “not yet measured” instead of guessing.
102 models
Model | Category | Parameters | Licence | Headline score |
|---|---|---|---|---|
| DeepSeek R1-0528 Qwen3 8B DeepSeek | Math and reasoning | 8.2 B | MIT | 86% AIME 2024, pass@1 |
| OpenReasoning Nemotron 32B NVIDIA | Math and reasoning | 32.8 B | CC-BY-4.0 | 89.2% AIME 2024, pass@1 |
| QwQ 32B Alibaba Qwen | Math and reasoning | 32.8 B | Apache-2.0 | 79.5% AIME 2024, pass@1 |
| Phi-4 reasoning Microsoft | Math and reasoning | 14.7 B | MIT | 75.3% AIME 2024, pass@1 |
| OpenMath Nemotron 7B NVIDIA | Math and reasoning | 7.6 B | CC-BY-4.0 | 74.8% AIME 2024, pass@1 |
| Phi-4 mini reasoning Microsoft | Math and reasoning | 3.8 B | MIT | 57.5% AIME 2024, pass@1 |
| DeepSeek R1 Distill Qwen 7B DeepSeek | Math and reasoning | 7.6 B | MIT | 55.5% AIME 2024, pass@1 |
| Devstral Small 1.1 Mistral AI | Coding | 23.6 B | Apache-2.0 | 53.6% SWE-bench Verified, % resolved |
| OpenHands LM 32B All Hands AI | Coding | 32.8 B | MIT | 37.2% SWE-bench Verified, % resolved |
| Qwen3 Coder 30B-A3B Alibaba Qwen | Coding | 30.5 B | Apache-2.0 | Not yet measured |
| Qwen2.5 Coder 32B Alibaba Qwen | Coding | 32.8 B | Apache-2.0 | Not yet measured |
| Qwen2.5 Coder 7B Alibaba Qwen | Coding | 7.6 B | Apache-2.0 | Not yet measured |
| DeepSeek Coder V2 Lite DeepSeek | Coding | 15.7 B | DeepSeek License | Not yet measured |
| Qwen2.5 VL 7B Alibaba Qwen | Vision-language | 8.3 B | Apache-2.0 | 58.6% MMMU (val) |
| Kimi-VL A3B Moonshot AI | Vision-language | 16.4 B | MIT | 57% MMMU (val) |
| SmolVLM2 2.2B Hugging Face | Vision-language | 2.2 B | Apache-2.0 | 42% MMMU (val) |
| Qwen3 VL 8B Alibaba Qwen | Vision-language | 8.8 B | Apache-2.0 | Not yet measured |
| Qwen3 VL 30B-A3B Alibaba Qwen | Vision-language | 31.1 B | Apache-2.0 | Not yet measured |
| InternVL3 8B Shanghai AI Lab | Vision-language | 7.9 B | Apache-2.0 | Not yet measured |
| Gemma 3 12B Google | Vision-language | 12.2 B | Gemma Terms of Use | Not yet measured |
| olmOCR 2 7B Ai2 | OCR and documents | 8.3 B | Apache-2.0 | 82.4% olmOCR-Bench, overall |
| PaddleOCR-VL Baidu PaddlePaddle | OCR and documents | 959 M | Apache-2.0 | 80% olmOCR-Bench, overall |
| DeepSeek-OCR DeepSeek | OCR and documents | 3.3 B | MIT | 75.7% olmOCR-Bench, overall |
| Nanonets OCR2 3B Nanonets | OCR and documents | 3.8 B | Not declared on Hugging Face Check terms before commercial use | 69.5% olmOCR-Bench, overall |
| dots.ocr rednote hilab | OCR and documents | 3.0 B | MIT | Not yet measured |
| Granite Docling 258M IBM | OCR and documents | 258 M | Apache-2.0 | Not yet measured |
| SmolDocling 256M IBM Research / Hugging Face | OCR and documents | 256 M | CDLA-Permissive-2.0 | Not yet measured |
| Canary-Qwen 2.5B NVIDIA | Speech to text | 2.6 B | CC-BY-4.0 | 6.37% Open ASR Leaderboard, average WER (English short-form) |
| Parakeet TDT 0.6B v2 NVIDIA | Speech to text | — | CC-BY-4.0 | 6.43% Open ASR Leaderboard, average WER (English short-form) |
| Parakeet TDT 0.6B v3 NVIDIA | Speech to text | 627 M | CC-BY-4.0 | 6.51% Open ASR Leaderboard, average WER (English short-form) |
| Distil-Whisper large-v3.5 Hugging Face | Speech to text | 756 M | MIT | 7.02% Open ASR Leaderboard, average WER (English short-form) |
| Voxtral Mini 3B Mistral AI | Speech to text | 4.7 B | Apache-2.0 | 7.84% Open ASR Leaderboard, average WER (English short-form) |
| Whisper large-v3 OpenAI | Speech to text | 1.5 B | Apache-2.0 | 7.84% Open ASR Leaderboard, average WER (English short-form) |
| Whisper large-v3 turbo OpenAI | Speech to text | 809 M | MIT | 8.39% Open ASR Leaderboard, average WER (English short-form) |
| Canary 1B Flash NVIDIA | Speech to text | 811 M | CC-BY-4.0 | Not yet measured |
| Kokoro 82M hexgrad | Text to speech | — | Apache-2.0 | Not yet measured |
| Chatterbox Resemble AI | Text to speech | — | MIT | Not yet measured |
| CSM 1B Sesame | Text to speech | 1.6 B | Apache-2.0 | Not yet measured |
| Dia 1.6B Nari Labs | Text to speech | 1.6 B | Apache-2.0 | Not yet measured |
| Orpheus 3B Canopy Labs | Text to speech | 3.8 B | Apache-2.0 | Not yet measured |
| VibeVoice 1.5B Microsoft | Text to speech | 2.7 B | MIT | Not yet measured |
| OpenAudio S1 mini Fish Audio | Text to speech | — | CC-BY-NC-SA-4.0 Check terms before commercial use | Not yet measured |
| HiDream-I1 Full HiDream.ai | Image generation | 17.1 B | MIT | 0.83 GenEval, overall |
| FLUX.1 [dev] Black Forest Labs | Image generation | 11.9 B | FLUX.1 [dev] Non-Commercial Check terms before commercial use | 0.66 GenEval, overall |
| Stable Diffusion XL Stability AI | Image generation | 2.6 B | CreativeML OpenRAIL++-M | 0.55 GenEval, overall |
| FLUX.1 [schnell] Black Forest Labs | Image generation | 11.9 B | Apache-2.0 | Not yet measured |
| FLUX.1 Kontext [dev] Black Forest Labs | Image generation | 11.9 B | FLUX.1 [dev] Non-Commercial Check terms before commercial use | Not yet measured |
| Qwen-Image Alibaba Qwen | Image generation | 20.4 B | Apache-2.0 | Not yet measured |
| Stable Diffusion 3.5 Medium Stability AI | Image generation | 2.5 B | Stability AI Community | Not yet measured |
| Qwen3 Embedding 8B Alibaba Qwen | Embeddings and rerankers | 7.6 B | Apache-2.0 | 70.58 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| Qwen3 Embedding 0.6B Alibaba Qwen | Embeddings and rerankers | 596 M | Apache-2.0 | 64.33 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| multilingual-e5-large-instruct Microsoft (intfloat) | Embeddings and rerankers | 560 M | MIT | 63.22 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| BGE-M3 BAAI | Embeddings and rerankers | — | MIT | 59.56 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| EmbeddingGemma 300M Google | Embeddings and rerankers | 303 M | Gemma Terms of Use | Not yet measured |
| Nomic Embed Text v1.5 Nomic AI | Embeddings and rerankers | 137 M | Apache-2.0 | Not yet measured |
| jina-embeddings-v3 Jina AI | Embeddings and rerankers | 572 M | CC-BY-NC-4.0 Check terms before commercial use | Not yet measured |
| Qwen3 Reranker 0.6B Alibaba Qwen | Embeddings and rerankers | 596 M | Apache-2.0 | 65.8 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| BGE Reranker v2 M3 BAAI | Embeddings and rerankers | 568 M | Apache-2.0 | 57.03 MMTEB (multilingual) mean task score; rerankers: MTEB-R |
| mxbai-rerank-large v2 Mixedbread | Embeddings and rerankers | 1.5 B | Apache-2.0 | Not yet measured |
| Qwen3 0.6B Alibaba Qwen | Small fast 'System 1' | 752 M | Apache-2.0 | 23.93% BFCL V4, overall accuracy |
| Llama 3.2 1B Meta | Small fast 'System 1' | 1.2 B | Llama 3.2 Community | 10.82% BFCL V4, overall accuracy |
| Gemma 3 1B Google | Small fast 'System 1' | 1.0 B | Gemma Terms of Use | 7.17% BFCL V4, overall accuracy |
| Gemma 3 270M Google | Small fast 'System 1' | 268 M | Gemma Terms of Use | Not yet measured |
| Qwen3 1.7B Alibaba Qwen | Small fast 'System 1' | 2.0 B | Apache-2.0 | Not yet measured |
| SmolLM3 3B Hugging Face | Small fast 'System 1' | 3.1 B | Apache-2.0 | Not yet measured |
| LFM2 1.2B Liquid AI | Small fast 'System 1' | 1.2 B | LFM Open License v1.0 | Not yet measured |
| Phi-4 mini Microsoft | Small fast 'System 1' | 3.8 B | MIT | Not yet measured |
| Granite 4.0 H Micro IBM | Small fast 'System 1' | 3.2 B | Apache-2.0 | Not yet measured |
| MedGemma 27B Google | Biomedical | 28.8 B | Health AI Developer Foundations terms | 85.3% MedQA (4-option), accuracy |
| MedGemma 1.5 4B Google | Biomedical | 4.3 B | Health AI Developer Foundations terms | 69.1% MedQA (4-option), accuracy |
| MedGemma 4B Google | Biomedical | 4.3 B | Health AI Developer Foundations terms | 64.4% MedQA (4-option), accuracy |
| MedGemma 27B text Google | Biomedical | 27.0 B | Health AI Developer Foundations terms | Not yet measured |
| TxGemma 9B chat Google | Biomedical | 9.2 B | Health AI Developer Foundations terms | Not yet measured |
| II-Medical 8B Intelligent Internet | Biomedical | 8.2 B | Apache-2.0 | Not yet measured |
| Med42 8B M42 | Biomedical | 8.0 B | Llama 3 Community | Not yet measured |
| OpenBioLLM 8B Saama AI (aaditya) | Biomedical | — | Llama 3 Community | Not yet measured |
| xLAM-2 8B Salesforce | Tool use and agents | 8.0 B | CC-BY-NC-4.0 Check terms before commercial use | 46.68% BFCL V4, overall accuracy |
| ToolACE-2 8B Team ACE | Tool use and agents | 8.0 B | Apache-2.0 | 42.44% BFCL V4, overall accuracy |
| Qwen3 4B Instruct 2507 Alibaba Qwen | Tool use and agents | 4.0 B | Apache-2.0 | 35.68% BFCL V4, overall accuracy |
| GLM-4.5 Air Z.ai | Tool use and agents | 110.5 B | MIT | Not yet measured |
| watt-tool 8B Watt AI | Tool use and agents | 8.0 B | Apache-2.0 | Not yet measured |
| Hermes 3 8B Nous Research | Tool use and agents | 8.0 B | Llama 3 Community | Not yet measured |
| Arch-Function 3B Katanemo | Tool use and agents | 3.1 B | Katanemo Research Check terms before commercial use | Not yet measured |
| Gemma 3 4B Google | Multilingual and translation | 4.3 B | Gemma Terms of Use | 1,273.2 LMArena text, non-English category (Elo-style) |
| Aya Expanse 32B Cohere Labs | Multilingual and translation | 32.3 B | CC-BY-NC-4.0 Check terms before commercial use | 1,213.4 LMArena text, non-English category (Elo-style) |
| Aya Expanse 8B Cohere Labs | Multilingual and translation | 8.0 B | CC-BY-NC-4.0 Check terms before commercial use | 1,180 LMArena text, non-English category (Elo-style) |
| Hunyuan-MT 7B Tencent | Multilingual and translation | 8.0 B | Not declared in HF metadata (see repo) Check terms before commercial use | Not yet measured |
| Seed-X PPO 7B ByteDance Seed | Multilingual and translation | 7.5 B | OpenMDW | Not yet measured |
| EuroLLM 9B UTTER project | Multilingual and translation | 9.2 B | Apache-2.0 | Not yet measured |
| Tower+ 9B Unbabel | Multilingual and translation | 9.2 B | CC-BY-NC-SA-4.0 Check terms before commercial use | Not yet measured |
| NLLB-200 600M Meta | Multilingual and translation | — | CC-BY-NC-4.0 Check terms before commercial use | Not yet measured |
| Chronos-2 Amazon | Time series | 119 M | Apache-2.0 | 0.698 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| TimesFM 2.5 Google | Time series | 231 M | Apache-2.0 | 0.705 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| Moirai 2.0 small Salesforce | Time series | 11 M | CC-BY-NC-4.0 Check terms before commercial use | 0.728 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| TimesFM 2.0 500M Google | Time series | 499 M | Apache-2.0 | 0.758 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| Chronos-Bolt base Amazon | Time series | 205 M | Apache-2.0 | 0.808 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| Chronos-Bolt small Amazon | Time series | 48 M | Apache-2.0 | 0.822 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| Granite TTM r2 IBM | Time series | 1 M | Apache-2.0 | 1.02 GIFT-Eval MASE, geometric mean relative to seasonal naive (lower is better) |
| SmolVLA Hugging Face LeRobot | Robotics (VLA) | 450 M | Apache-2.0 | 87.3% LIBERO, average success rate (4 suites) |
| π0 (LeRobot port) Physical Intelligence | Robotics (VLA) | 3.5 B | Apache-2.0 | 86% LIBERO, average success rate (4 suites) |
| OpenVLA 7B Stanford / OpenVLA | Robotics (VLA) | 7.5 B | MIT | 76.5% LIBERO, average success rate (4 suites) |
| GR00T N1.5 3B NVIDIA | Robotics (VLA) | 2.7 B | NVIDIA One-Way Non-Commercial Check terms before commercial use | Not yet measured |
As JSON: /api/v1/specialists.json.