Data as of · 78 models, 80 questions · reviewed by LLM Verifications team · methodology

Open weights · tested 27 Oct 2026

Best open-weights AI models

45 of the 78 models we test have published weights you can download and run yourself. Here they are on the same benchmarks as every other model. The best of them, Muse Spark Open, scores 77.5%, against 85.0% for Claude Opus 5.5, the top model without listed open weights.

Score vs. cost per question

Open-weights models in color, models without listed open weights in grey

Most attractive quadrant0255075100$0.00001$0.0001$0.001$0.01$0.10Cost per question (USD, log) →Verification score (%) →Qwen3.8 2.4T A95BCommand AReka Flash 3Gemma 4 26B A4BGemma 4 31BQwen3.5 397B A17BHermes 4 405BMuse WonderGrok 4.7o3gpt-oss-20bGranite 4.0 MicroMinistral 3 3B 2512Nova Micro 1.0Mercury 2.5Phi 4Granite 4.2 8BHy3GLM 5.3 FlashSolar Pro 4MiniMax M3Mistral Large 4Claude Opus 5.5DeepSeek V4.1 Flash 1017Gemini 4 ArgonGPT-5.4Llama 4 Maverick

All open-weights models

Leaderboard
#
1Muse Spark Open
Meta · served by Together
77.562 / 802200.5 s$0.0045$1.25 / $4.251M
2DeepSeek V4.1 Flash 1017
DeepSeek · served by Wafer
73.859 / 80570.9 s$0.00091$0.15 / $0.61M
3DeepSeek V4 Flash Vision Exp
DeepSeek · served by DeepInfra
71.257 / 801700.7 s$0.0011$0.2156 / $0.64681M
4DeepSeek V4 Pro 0813
DeepSeek · served by Alibaba
71.257 / 80560.8 s$0.0036$1.32 / $3.961M
5MiMo-V2.6-Pro
Xiaomi · served by DeepInfra
68.855 / 80580.7 s$0.002$0.435 / $0.871.1M
6GLM 5.3 Flash
Z.ai · served by Wafer
67.554 / 80830.5 s$0.00015$0.15 / $0.51M
7Inkling Small
Thinking Machines · served by DeepInfra
67.554 / 802330.5 s$0.00047$0.45 / $1.2524k
8GLM 5.3
Z.ai · served by Wafer
67.554 / 801540.8 s$0.00093$0.079 / $1.41M
9Inkling
Thinking Machines · served by Together
66.253 / 801180.6 s$0.0012$1 / $4.05524k
10MiniMax M3.1 Flash
MiniMax · served by InferenceNet
66.253 / 801310.5 s$0.00087$0.15 / $0.61M
11Kimi K2.7 Code
MoonshotAI · served by Inceptron
65.052 / 801360.4 s$0.006$0.56 / $3.1262k
12Step 5 Preview
StepFun · served by InferenceNet
63.851 / 80930.6 s$0.0047$1 / $2.71M
13Qwen3.8 27B
Qwen · served by Modal
63.851 / 80490.5 s$0.0043$0.2 / $2.51M
14MiMo-V2.6-Flash
Xiaomi · served by Relace
62.550 / 80770.4 s$0.00056$0.14 / $0.281.1M
15Qwen3 Coder Next
Qwen · served by Parasail
62.550 / 801060.4 s$0.0017$0.12 / $0.8262k
16Nemotron 3 Ultra
NVIDIA · served by BaseTen
60.048 / 801360.8 s$0.0057$0.6 / $2.4262k
17Llama 4 Maverick
Meta · served by DigitalOcean
59.046 / 78130.7 s$0.00059$0.1875 / $0.6525524k
18Hy3
Tencent · served by Tencent
58.847 / 80961.7 s$0.0012$0.132 / $0.528262k
19gpt-oss-120b
OpenAI · served by CoreWeave
58.847 / 80820.5 s$0.00013$0.15 / $0.6131k
20Qwen3.8 2.4T A95B
Qwen · served by Venice
57.546 / 80443.1 s$0.015$2 / $61M
21MiniMax M3
MiniMax · served by CoreWeave
57.546 / 801210.5 s$0.0026$0.3 / $1.21M
22gpt-oss-20b
OpenAI · served by Darkbloom
51.241 / 801030.7 s$0.000066$0.018 / $0.09131k
23LongCat 2.0
Meituan · served by AtlasCloud
48.839 / 80671.4 s$0.0032$0.3 / $1.21M
24Command A
Cohere · served by Cohere
45.036 / 8071.8 s$0.007$2.5 / $10256k
25Nemotron 3.5 Lightning
NVIDIA · served by Io Net
43.835 / 801160.5 s$0.00049$0.07 / $0.2262k
26Ministral 3 8B 2512
Mistral · served by Mistral
42.534 / 801240.3 s$0.00027$0.15 / $0.15262k
27Llama 3.3 70B Instruct
Meta · served by DeepInfra
42.534 / 80420.7 s$0.00053$0.1 / $0.32131k
28Llama 4 Scout
Meta · served by DeepInfra
41.233 / 80151.2 s$0.00037$0.1 / $0.31.3M
29Phi 4
Microsoft · served by DeepInfra
41.233 / 801030.5 s$0.00012$0.07 / $0.1416k
30Nemotron 3.5 Super VL
NVIDIA · served by Darkbloom
40.032 / 807150.6 s$0.0013$0.15 / $0.51M
31Qwen3.6 35B A3B
Qwen · served by DeepInfra
37.530 / 801350.4 s$0.0037$0.1 / $0.95262k
32Granite 4.2 8B
IBM · served by DeepInfra
36.229 / 80600.6 s$0.000077$0.06 / $0.25131k
33Ministral 3 14B 2512
Mistral · served by Mistral
33.827 / 80320.6 s$0.00031$0.2 / $0.2262k
34Qwen3.5 397B A17B
Qwen · served by Phala
32.526 / 802400.6 s$0.013$0.5 / $3.6262k
35Mistral Small 4
Mistral · served by Mistral
32.526 / 801700.5 s$0.0019$0.15 / $0.6262k
36Gemma 4 31B
Google · served by CoreWeave
30.024 / 80580.6 s$0.0014$0.2 / $0.4262k
37Hermes 4 405B
Nous · served by Nebius
26.221 / 80410.3 s$0.01$1 / $3131k
38Ministral 3 3B 2512
Mistral · served by Mistral
26.221 / 802520.3 s$0.00019$0.1 / $0.1131k
39Granite 4.0 Micro
IBM · served by Cloudflare
25.020 / 80380.7 s$0.000051$0.017 / $0.112131k
40Gemma 4 26B A4B
Google · served by Darkbloom
15.012 / 801720.7 s$0.0011$0.09 / $0.3262k
41Reka Flash 3
rekaai · served by Reka
1.21 / 801010.5 s$0.0008$0.1 / $0.266k
42LFM2.5-2.6B—0 / 0———$0.35 / $0.5128k
43Ling 3.0 Flash VL—0 / 0———$0.06 / $0.18131k
44Qwen3.5-122B-A10B
Qwen · served by SiliconFlow
—0 / 01081.2 s—$0.4 / $3.2262k
45Qwen3 Next 80B A3B Thinking—0 / 0———$0.15 / $1.2262k

Frequently asked questions

What is the best open-weights AI model?
Muse Spark Open by Meta has the highest verification score of the open-weights models we test: 77.5% (62 of 80 questions correct).
How close are open-weights models to the best closed models?
The best open-weights model, Muse Spark Open, scored 77.5%; the best model without listed open weights, Claude Opus 5.5, scored 85.0%. That 7.5-point gap is within what chance could produce on our 80 questions.
What is the fastest open-weights model?
Nemotron 3.5 Super VL, at a median 715 output tokens per second in our runs.
What is the cheapest open-weights model?
Granite 4.0 Micro, at $0.000051 per question as billed.
How do you decide a model is open-weights?
A model counts as open-weights when OpenRouter's catalogue links it to a published Hugging Face repository. Licenses vary, and some restrict commercial use, so check the repository before you deploy.