Model comparison
Compare open-weight AI models
This page compares 59 open-weight AI models, from text and code to vision, speech and embeddings. Benchmark results come from Epoch AI and LMArena, model details from Hugging Face, and every figure names its source and the day it was read.
Go to the table
Data as of Sep 24, 2026
Filter and sort
Open-weight AI models with public benchmark results and model details. Select up to four rows to compare them side by side. Scroll sideways to see every column.
| Compare | Modality | Licence | Origin | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Kimi K3moonshotai |
| kimi-k3Publisher’s terms | 2.8T | 1,485 | 93.1% | 97.2% | No value published by the source for this model | 50.6% | No value published by the source for this model | 1,660 | No value published by the source for this model | 1M | 1.8M | China | |
| GLM-5.3-Flashzai-org |
| mitPermissive | 321.3B | 1,475 | 90.2% | 93.9% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,612 | 1,299 | 1M | 4M | China | |
| GLM-5.2zai-org |
| mitPermissive | 753.3B | 1,472 | 91.9% | 86.4% | 78.7% | 34.2% | No value published by the source for this model | 1,600 | No value published by the source for this model | 1M | 927.7K | China | |
| MiMo-V2.5-ProXiaomiMiMo |
| mitPermissive | 1T | 1,467 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,476 | No value published by the source for this model | 1M | 22.5K | China | |
| GLM-5.1zai-org |
| mitPermissive | 753.9B | 1,466 | 89.9% | 93.3% | 74.2% | 34.0% | No value published by the source for this model | 1,508 | No value published by the source for this model | 202.8K | 74K | China | |
| Kimi K2.6moonshotai |
| modified-mitPublisher’s terms | 1T | 1,460 | 90.8% | 96.1% | 76.7% | 34.9% | No value published by the source for this model | 1,509 | 1,280 | 262.1K | 435K | China | |
| DeepSeek-V4-Prodeepseek-ai |
| mitPermissive | 1.6T | 1,457 | 90.9% | 96.7% | 77.6% | 47.0% | No value published by the source for this model | 1,445 | No value published by the source for this model | 1M | 514.4K | China | |
| Hy3tencent |
| apache-2.0Permissive | 298.8B | 1,456 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,509 | No value published by the source for this model | 262.1K | 12.8K | No value published by the source for this model | |
| Gemma 4 31Bgoogle |
| apache-2.0Permissive | 31.3B | 1,451 | 75.8% | 73.3% | No value published by the source for this model | 10.4% | No value published by the source for this model | 1,364 | 1,276 | 262.1K | 9.3M | United States | |
| Qwen3.5-397B-A17BQwen |
| apache-2.0Permissive | 403.4B | 1,442 | 86.4% | 88.9% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,399 | 1,265 | 262.1K | 199.7K | China | |
| GLM-4.7zai-org |
| mitPermissive | 358.3B | 1,442 | 83.3% | 83.3% | No value published by the source for this model | 32.2% | No value published by the source for this model | 1,435 | No value published by the source for this model | 202.8K | 101.1K | China | |
| MiniMax-M3MiniMaxAI |
| minimax-communityPublisher’s terms | 427B | 1,441 | 90.9% | 71.1% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,485 | 1,254 | 1M | 167.8K | China | |
| Inklingthinkingmachines |
| apache-2.0Permissive | 952.4B | 1,440 | 88.3% | 88.9% | No value published by the source for this model | 40.3% | No value published by the source for this model | 1,412 | No value published by the source for this model | No value published by the source for this model | 369.2K | United States | |
| Gemma 4 26B A4Bgoogle |
| apache-2.0Permissive | 25.8B | 1,438 | 73.2% | 82.2% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,362 | 1,259 | 262.1K | 10.7M | United States | |
| Qwen3.8-27BQwen |
| apache-2.0Permissive | 27.8B | 1,437 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,591 | 1,272 | 262.1K | 6.8M | China | |
| DeepSeek-V4-Flashdeepseek-ai |
| mitPermissive | 290.9B | 1,436 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1M | 1.4M | China | |
| Mistral Medium 3.5mistralai |
| otherPublisher’s terms | 127.7B | 1,426 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,265 | 1,222 | 262.1K | 99.2K | No value published by the source for this model | |
| DeepSeek-V3.2deepseek-ai |
| mitPermissive | 685.4B | 1,425 | 83.4% | 87.8% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,325 | No value published by the source for this model | 163.8K | 2.8M | China | |
| DeepSeek-R1-0528deepseek-ai |
| mitPermissive | 684.5B | 1,421 | 76.3% | 66.4% | No value published by the source for this model | No value published by the source for this model | 96.6% | No value published by the source for this model | No value published by the source for this model | 163.8K | 157.4K | China | |
| Qwen3.5-122B-A10BQwen |
| apache-2.0Permissive | 125.1B | 1,417 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,358 | 1,245 | 262.1K | 323.7K | China | |
| Qwen3-VL-235B-A22BQwen |
| apache-2.0Permissive | 235.7B | 1,414 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,247 | 262.1K | 329.6K | No value published by the source for this model | |
| Mistral Large 3mistralai |
| apache-2.0Permissive | No value published by the source for this model | 1,413 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,229 | 1,223 | No value published by the source for this model | 2.2K | No value published by the source for this model | |
| Inkling-Smallthinkingmachines |
| apache-2.0Permissive | 266B | 1,405 | 88.5% | 90.0% | No value published by the source for this model | 19.1% | No value published by the source for this model | 1,407 | 1,236 | No value published by the source for this model | 449.6K | United States | |
| Qwen3-235B-A22B-Thinking-2507Qwen |
| apache-2.0Permissive | 235.1B | 1,400 | 80.1% | 86.7% | No value published by the source for this model | 40.4% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 262.1K | 13.7K | China | |
| Qwen3-Coder-480B-A35BQwen |
| apache-2.0Permissive | 480.2B | 1,388 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,274 | No value published by the source for this model | 262.1K | 35.1K | China | |
| Qwen3-30B-A3B-Instruct-2507Qwen |
| apache-2.0Permissive | 30.5B | 1,382 | 55.6% | 62.2% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 262.1K | 789K | China | |
| GLM-4.6Vzai-org |
| mitPermissive | 107.7B | 1,378 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,163 | 131.1K | 3.7K | No value published by the source for this model | |
| Gemma 3 27BgoogleGated |
| gemmaPublisher’s terms | 27.4B | 1,365 | 47.7% | 22.5% | No value published by the source for this model | No value published by the source for this model | 74.0% | No value published by the source for this model | 1,164 | No value published by the source for this model | 389.6K | United States | |
| Mistral Small 3.2mistralai |
| apache-2.0Permissive | 24B | 1,357 | 49.1% | 30.3% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1,141 | 131.1K | 199.7K | France | |
| Command ACohereLabsGated |
| cc-by-nc-4.0Non-commercial | 111.1B | 1,354 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 2.5K | No value published by the source for this model | |
| gpt-oss-120bopenai |
| apache-2.0Permissive | 116.8B | 1,352 | 75.8% | 88.9% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 131.1K | 4.8M | United States | |
| Qwen3-32BQwen |
| apache-2.0Permissive | 32.8B | 1,347 | 65.7% | 66.9% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 41K | 5.1M | China | |
| Olmo 3.1 32B Instructallenai |
| apache-2.0Permissive | 32.2B | 1,330 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 65.5K | 15.6K | No value published by the source for this model | |
| Llama 4 Maverickmeta-llamaGated |
| llama4Publisher’s terms | 401.6B | 1,327 | 67.0% | 20.6% | No value published by the source for this model | No value published by the source for this model | 73.0% | No value published by the source for this model | 1,142 | No value published by the source for this model | 10K | United States | |
| Llama 4 Scoutmeta-llamaGated |
| llama4Publisher’s terms | 108.6B | 1,321 | 51.8% | 7.8% | No value published by the source for this model | No value published by the source for this model | 62.3% | No value published by the source for this model | 1,118 | No value published by the source for this model | 161.8K | United States | |
| Llama 3.3 70Bmeta-llamaGated |
| llama3.3Publisher’s terms | 70.6B | 1,318 | 47.4% | 5.1% | No value published by the source for this model | No value published by the source for this model | 41.6% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 916.8K | United States | |
| gpt-oss-20bopenai |
| apache-2.0Permissive | 20.9B | 1,317 | 60.8% | 65.3% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 131.1K | 6.7M | United States | |
| Command R+ (08-2024)CohereLabsGated |
| cc-by-nc-4.0Non-commercial | 103.8B | 1,276 | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 142 | No value published by the source for this model | |
| Phi-4microsoft |
| mitPermissive | 14.7B | 1,256 | 56.1% | 13.8% | No value published by the source for this model | No value published by the source for this model | 64.9% | No value published by the source for this model | No value published by the source for this model | 16.4K | 621.5K | United States | |
| Mixtral 8x22Bmistralai |
| apache-2.0Permissive | 140.6B | 1,229 | 34.1% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 24.2% | No value published by the source for this model | No value published by the source for this model | 65.5K | 30.3K | France | |
| Llama 3.1 8Bmeta-llamaGated |
| llama3.1Publisher’s terms | 8B | 1,211 | 27.0% | 1.7% | No value published by the source for this model | No value published by the source for this model | 22.9% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 6.2M | United States | |
| BGE-M3BAAI |
| mitPermissive | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 8.2K | 37.2M | No value published by the source for this model | |
| ChatterboxResembleAI |
| mitPermissive | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 1.8M | No value published by the source for this model | |
| Devstral Small 2mistralai |
| apache-2.0Permissive | 24B | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 393.2K | 305.2K | France | |
| Granite 4.1 30Bibm-granite |
| apache-2.0Permissive | 28.9B | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 131.1K | 550.2K | United States | |
| Kimi K2.7 Codemoonshotai |
| modified-mitPublisher’s terms | 1T | No value published by the source for this model | 87.9% | 95.6% | No value published by the source for this model | 36.5% | No value published by the source for this model | 1,472 | No value published by the source for this model | 262.1K | 104.7K | China | |
| Kokoro-82Mhexgrad |
| apache-2.0Permissive | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 11.9M | No value published by the source for this model | |
| Magistral Small 1.2mistralai |
| apache-2.0Permissive | 24B | No value published by the source for this model | 47.6% | 28.1% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 131.1K | 13.6K | France | |
| Mistral 7B v0.3mistralai |
| apache-2.0Permissive | 7.2B | No value published by the source for this model | 15.2% | 0.3% | No value published by the source for this model | No value published by the source for this model | 3.7% | No value published by the source for this model | No value published by the source for this model | 32.8K | 2.3M | France | |
| multilingual-e5-largeintfloat |
| mitPermissive | 560M | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 514 | 7.3M | No value published by the source for this model | |
| Nemotron 3 Ultranvidia |
| openmdw-1.1Publisher’s terms | 560.5B | No value published by the source for this model | 85.4% | 86.7% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 262.1K | 171.3K | United States | |
| nomic-embed-text-v1.5nomic-ai |
| apache-2.0Permissive | 137M | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 2K | 14.2M | No value published by the source for this model | |
| Parakeet TDT 0.6B v3nvidia |
| cc-by-4.0Permissive | 627M | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 571.6K | No value published by the source for this model | |
| Qwen3-Embedding-8BQwen |
| apache-2.0Permissive | 7.6B | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 41K | 2.6M | No value published by the source for this model | |
| Qwen3.6-27BQwen |
| apache-2.0Permissive | 27.8B | No value published by the source for this model | 85.9% | 91.1% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 262.1K | 3M | China | |
| Qwen3.6-35B-A3BQwen |
| apache-2.0Permissive | 36B | No value published by the source for this model | 84.8% | 86.7% | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 262.1K | 3.1M | China | |
| Voxtral Smallmistralai |
| apache-2.0Permissive | 24.3B | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 131.1K | 24.4K | No value published by the source for this model | |
| Whisper large-v3openai |
| apache-2.0Permissive | 1.5B | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 4.7M | No value published by the source for this model | |
| Whisper large-v3-turboopenai |
| mitPermissive | 809M | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | No value published by the source for this model | 6.5M | No value published by the source for this model |
Side by side
Compare selected models
Select two to four models in the table to compare them here.
What the figures measure
The benchmarks, in plain words
Each benchmark tests one narrow skill under fixed conditions. Read them together, check the date, and test on your own tasks before choosing a model.
- LMArena Text
Human preference on general chat. People compare two anonymous answers to the same prompt and vote; LMArena turns the votes into a rating (Bradley–Terry, on an Elo-like scale) with style control on, so longer or more formatted answers are not rewarded for form alone. Compare ratings together with their interval, not ranks alone.
- GPQA Diamond
Graduate-level multiple-choice questions in biology, physics and chemistry, written by domain experts to be hard to look up. Epoch AI’s own runs; the score is the share answered correctly, where random guessing scores a quarter.
- OTIS Mock AIME 2024–2025
Competition mathematics: problems from mock AIME exams written for the OTIS programme, with a single numeric answer each. Epoch AI’s own runs; share of problems solved.
- SWE-bench Verified
Real software engineering: fixing actual GitHub issues in open-source Python repositories, checked by the projects’ tests, on a subset verified by human engineers. Epoch AI’s own runs; share of issues resolved.
- SimpleQA Verified
Short factual questions with one verifiable answer, a test of what a model knows without search and of how often it makes things up. Epoch AI’s own runs; share answered correctly.
- MATH Level 5
The hardest level of the MATH dataset of competition problems. Mostly run on older models, so newer ones often show “—”. Epoch AI’s own runs; share solved.
- LMArena WebDev
Human preference on building web apps. Voters compare two models’ working web pages for the same request; the rating uses the same method as the text arena.
- LMArena Vision
Human preference on prompts that include an image, such as charts, photos or documents; same rating method as the text arena.
Sources
Sources, licences and freshness
Azinove reads these public sources on its own server, at most once a day, and keeps a saved copy so the page keeps working when a source is unavailable.
Hugging Face Hub
Last read on Sep 24, 2026
- Used for
- Parameters, licence, context, downloads, likes and last update of each repository
- Citation
- Model metadata from the Hugging Face Hub API (huggingface.co), each row linked to its repository.
Epoch AI
Last read on Sep 24, 2026
- Used for
- GPQA Diamond, OTIS Mock AIME 2024–2025, SWE-bench Verified, SimpleQA Verified, MATH Level 5, plus organisation and country
- Licence
- CC BY 4.0
- Citation
- Epoch AI, ‘Capabilities & Benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource].
LMArena
Last read on Sep 22, 2026
- Used for
- LMArena Text, LMArena WebDev, LMArena Vision, from the public leaderboard dataset
- Licence
- CC BY 4.0
- Citation
- LMArena, Arena Leaderboard Dataset (lmarena-ai/leaderboard-dataset), Hugging Face, CC BY 4.0.
Azinove AI early access
Tell us which models you need
Azinove AI runs open-weight models on GPUs Azinove operates. During early access, the models, hosting location and terms are agreed with each company in writing, so the models your team needs shape what is offered first.