Score73.6
License
MIT
OSI-approved
Parameters
1020B MoE (42B active)
FP8, Q4
Context1M
Reference hardware
8× NVIDIA H200 (141GB) · 1128GB total
Recommendation
Ranking data as of
Open-Source LLM Leaderboard 2026
The best open-source LLM by current open-weight benchmark score is MiMo-V2.6-Pro. It leads the October 2026 open-weight ranking at 73.6, ahead of Qwen3.8 Max (69.7) and Step 5 Preview (69.4). The license directory below separates OSI-approved licenses from community terms.
Bottom line: use the live table above for capability order, then check evidence status, license terms, memory requirements, and serving cost before choosing a deployment.
Citable stat102 open-weight models are ranked as of October 17, 2026; MiMo-V2.6-Pro leads at 73.6/100.
Full Rankings (102 ranked, 1 tracked)
4
Ling 3.1 Flash
InclusionAI·Open Weight·262K
68
BenchAlign v5.8
Estimated
Conditional range 53.68–82.38
6
DeepSeek V4.1 Flash 1017
DeepSeek·Open Weight·1M
65.6
BenchAlign v5.8
Supported
90% interval 60.59–70.63
8
dots3-note Preview
Dots Studio·Open Weight·512K
63.8
BenchAlign v5.8
Estimated
Conditional range 49.43–78.13
9
DeepSeek V4 Pro 0813
DeepSeek·Open Weight·1M
63.2
BenchAlign v5.8
Estimated
Conditional range 48.84–77.54
10
Qwen3.8-Flash-Next
Alibaba·Open Weight·262K
62.8
BenchAlign v5.8
Estimated
Conditional range 48.45–77.15
12
Ornith-1.5-397B
Ornith AI·Open Weight·262K
62.3
BenchAlign v5.8
Estimated
Conditional range 47.98–76.68
17
Kimi K2.6
Moonshot AI·Open Weight·256K
58.4
BenchAlign v5.8
Estimated
Conditional range 48.68–68.13
19
Qwen3.5 397B
Alibaba·Open Weight·128K
55.3
BenchAlign v5.8
Estimated
Conditional range 40.99–69.69
22
Agents-A1
InternScience·Open Weight·262K
54.2
BenchAlign v5.8
Estimated
Conditional range 39.85–68.55
23
Agents-A1-F16-GGUF
InternScience·Open Weight·262K
54.2
BenchAlign v5.8
Estimated
Conditional range 39.85–68.55
24
Agents-A1-FP8
InternScience·Open Weight·262K
54.2
BenchAlign v5.8
Estimated
Conditional range 39.85–68.55
25
Agents-A1-Q4_K_M-GGUF
InternScience·Open Weight·262K
54.2
BenchAlign v5.8
Estimated
Conditional range 39.85–68.55
26
Agents-A1-Q8_0-GGUF
InternScience·Open Weight·262K
54.2
BenchAlign v5.8
Estimated
Conditional range 39.85–68.55
27
Ternary Bonsai 2 27B
Prism ML·Open Weight·262K
53.9
BenchAlign v5.8
Estimated
Conditional range 39.52–68.22
28
Inkling-Small
Thinking Machines Lab·Open Weight·1M
53.5
BenchAlign v5.8
Supported
90% interval 45.05–61.94
30
Inkling
Thinking Machines Lab·Open Weight·1M
52.6
BenchAlign v5.8
Supported
90% interval 42.27–62.94
32
Hy3 Preview
Tencent·Open Weight·256K
52.2
BenchAlign v5.8
Estimated
Conditional range 37.86–66.56
34
Kimi K2.7 Code
Moonshot AI·Open Weight·256K
49.5
BenchAlign v5.8
Supported
90% interval 42.30–56.74
35
Nemotron 3.5 Super VL 120B A12B
NVIDIA·Open Weight·1M
49.4
BenchAlign v5.8
Estimated
Conditional range 35.04–63.74
37
Qwen3.6-27B
Alibaba·Open Weight·262K
48.6
BenchAlign v5.8
Estimated
Conditional range 34.24–62.94
38
Qwen3.5-27B
Alibaba·Open Weight·262K
48.4
BenchAlign v5.8
Estimated
Conditional range 38.68–58.13
40
Ling 3.0 Flash VL
InclusionAI·Open Weight·262K
46.7
BenchAlign v5.8
Estimated
Conditional range 32.37–61.07
41
Nemotron 3 Ultra
NVIDIA·Open Weight·1M
46.6
BenchAlign v5.8
Estimated
Conditional range 32.24–60.93
44
Nemotron 3.5 Bolt 74B-A9B
NVIDIA·Open Weight·256K
45.2
BenchAlign v5.8
Estimated
Conditional range 30.86–59.56
45
Ling 3.0 Flash
InclusionAI·Open Weight·262K
45.1
BenchAlign v5.8
Estimated
Conditional range 30.76–59.46
46
Step 3.7 Flash
StepFun·Open Weight·256K
43.7
BenchAlign v5.8
Estimated
Conditional range 29.30–58.00
47
Qwen3.6-35B-A3B
Alibaba·Open Weight·262K
43.5
BenchAlign v5.8
Estimated
Conditional range 29.16–57.86
48
Step 3.5 Flash
StepFun·Open Weight·256K
42.1
BenchAlign v5.8
Estimated
Conditional range 32.39–51.84
49
Qwen3.5-122B-A10B
Alibaba·Open Weight·262K
41.7
BenchAlign v5.8
Estimated
Conditional range 32.02–51.46
50
Muse Glimmer 30B
Meta·Open Weight·131K
41.2
BenchAlign v5.8
Estimated
Conditional range 26.84–55.54
56
DeepSeek V3.1 (Reasoning)
DeepSeek·Open Weight·128K
39.2
BenchAlign v5.8
Estimated
Conditional range 29.46–48.90
58
LFM2.5-24B-A2B
LiquidAI·Open Weight·131K
37.9
BenchAlign v5.8
Estimated
Conditional range 23.55–52.25
60
DeepSeek V3 0324
DeepSeek·Open Weight·128K
36.5
BenchAlign v5.8
Supported
90% interval 20.08–52.91
63
Mistral Small 4
Mistral·Open Weight·256K
34.7
BenchAlign v5.8
Estimated
Conditional range 24.98–44.43
64
Trinity-Large-Thinking
Arcee AI·Open Weight·512K
34.5
BenchAlign v5.8
Estimated
Conditional range 24.78–44.23
66
Ornith-1.5-35B-A3B
Ornith AI·Open Weight·262K
34
BenchAlign v5.8
Estimated
Conditional range 19.63–48.33
67
Mistral Medium 3.5 128B
Mistral·Open Weight·256K
33.6
BenchAlign v5.8
Estimated
Conditional range 19.27–47.97
70
Ling 2.6 Flash
InclusionAI·Open Weight·262K
32.4
BenchAlign v5.8
Estimated
Conditional range 18.08–46.78
71
Nemotron 3 Nano Omni 30B A3B
NVIDIA·Open Weight·256K
32.2
BenchAlign v5.8
Estimated
Conditional range 17.87–46.57
74
Ornith-1.5-9B
Ornith AI·Open Weight·262K
31.3
BenchAlign v5.8
Estimated
Conditional range 16.98–45.67
75
LFM2.5-2.6B
LiquidAI·Open Weight·128K
31.2
BenchAlign v5.8
Estimated
Conditional range 16.85–45.55
77
DeepSeek R1 Distill Qwen 32B
DeepSeek·Open Weight·128K
30.8
BenchAlign v5.8
Estimated
Conditional range 16.46–45.16
78
LFM2.5-8B-A1B
LiquidAI·Open Weight·128K
30.1
BenchAlign v5.8
Estimated
Conditional range 15.77–44.47
80
Granite-4.0-H-1B
IBM·Open Weight·128K
28.8
BenchAlign v5.8
Estimated
Conditional range 14.46–43.16
81
Qwen2.5-72B
Alibaba·Open Weight·128K
28.7
BenchAlign v5.8
Estimated
Conditional range 19.02–38.46
84
Granite-4.0-350M
IBM·Open Weight·32K
28.5
BenchAlign v5.8
Estimated
Conditional range 14.10–42.80
85
Granite-4.0-H-350M
IBM·Open Weight·32K
28.5
BenchAlign v5.8
Estimated
Conditional range 14.10–42.80
86
Exaone 4.0 32B
LG AI Research·Open Weight·128K
28.4
BenchAlign v5.8
Estimated
Conditional range 14.04–42.74
87
Llama 3.1 405B
Meta·Open Weight·128K
28.2
BenchAlign v5.8
Estimated
Conditional range 18.48–37.93
89
Qwen3-Omni-30B-A3B-Instruct
Alibaba·Open Weight·N/A
28.1
BenchAlign v5.8
Estimated
Conditional range 13.73–42.43
90
Exaone 4.0 1.2B
LG AI Research·Open Weight·128K
27.4
BenchAlign v5.8
Estimated
Conditional range 13.00–41.70
91
Qwen2.5 Coder 32B Instruct
Alibaba·Open Weight·128K
26.6
BenchAlign v5.8
Estimated
Conditional range 16.89–36.33
92
Nemotron 3.5 Lightning 30B A3B NVFP4
NVIDIA·Open Weight·1M
25.5
BenchAlign v5.8
Estimated
Conditional range 11.15–39.85
94
Ministral 3 14B
Mistral·Open Weight·128K
24.9
BenchAlign v5.8
Estimated
Conditional range 15.22–34.66
97
Ministral 3 8B
Mistral·Open Weight·128K
20
BenchAlign v5.8
Estimated
Conditional range 10.32–29.77
98
Ministral 3 3B
Mistral·Open Weight·128K
18.9
BenchAlign v5.8
Estimated
Conditional range 9.16–28.60
99
Mixtral 8x22B Instruct v0.1
Mistral·Open Weight·64K
18.2
BenchAlign v5.8
Estimated
Conditional range 8.44–27.89
100
Nemotron-4 15B
NVIDIA·Open Weight·32K
9.9
BenchAlign v5.8
Estimated
Conditional range 0.21–19.66
101
Mistral 7B v0.3
Mistral·Open Weight·32K
7.2
BenchAlign v5.8
Estimated
Conditional range 0.00–16.93
—
Mistral Large 4
Mistral·Pending·1M
Public preview through Mistral Studio; Mistral says the weights are due by the end of October 2026.
—
tracked
Current position
MiMo-V2.6-Pro leads the live open-weight ranking at 73.6 with Estimated evidence.
Qwen3.8 Max ranks #2 at 69.7 with Supported evidence.
Step 5 Preview ranks #3 at 69.4 with Estimated evidence.
How to choose
Best open-weight model overall?The #1 row of the ranking above, then its evidence label and license termsCoding-focused open model?Use the open rows on the calibrated coding leaderboardSelf-hosting economics?Estimate hardware and API break-even before choosing a model familyChinese-lab open models?See the Chinese model rankings — most frontier open weights come from these labs
Key Takeaways
The top model is MiMo-V2.6-Pro by Xiaomi with a BenchAlign v5.8 score of 73.6 and Estimated evidence.
The best open-weight model is MiMo-V2.6-Pro at position #1.
102 models are ranked here; 1 tracked model is listed without a position.
Score in Context
What these scores mean
Open-weight models are ranked by the same public BenchAlign v5.8 overall score as proprietary models. Supported and Estimated describe the evidence behind each position; they are not separate leaderboards.
Known limitations
Open weight is not the same as OSI-approved open source. The ranking does not score license restrictions, memory requirements, serving throughput, fine-tuning support, or the engineering cost of operating the model.
About this ranking
Ranking data as of October 17, 2026
This page is the canonical open-weight ranking. It uses the same public BenchAlign v5.8 overall lane as the main leaderboard, then filters to downloadable model weights. The score compares measured capability; it does not decide whether a license is permissive, a model fits your hardware, or self-hosting beats an API on cost. Use the linked decision guide for those deployment questions.
This is the public BenchAlign v5.8 overall lane filtered to open-weight models. The table orders point estimates. Evidence labels describe source support; conditional ranges do not establish rank confidence.
The open-weight slice starts with MiMo-V2.6-Pro, followed by Qwen3.8 Max and Step 5 Preview. All rows use the public BenchAlign v5.8 projection contract. Evidence badges and score intervals matter as much as small point gaps because public source coverage is uneven.
Every ranked model publishes downloadable weights, but that does not make every license OSI-approved or every deployment practical. Use the directory below to check license terms, quantization, and reference hardware before shortlisting a model.
This ranking uses the public BenchAlign v5.8 overall contract filtered to open-weight models. For detailed model profiles, click any model name above. To compare two specific models head-to-head, use the "vs #" links.
Deployment evidence
Compare license, size, and deployment
The performance table ranks every eligible open-weight row. This smaller directory includes only models with a deployment record in the self-host catalog, so license and hardware claims remain auditable instead of being filled from family names.
Open weight is an access category, not a license verdict. “OSI-approved” appears only when the deployment catalog records MIT or Apache 2.0. Community and custom licenses stay separate, even when their weights are downloadable.
Showing 40 of 40 deployment-documented rows
Deployment catalog checked 2026-10-16.
- Score69.4License
StepFun Model License 1.0
Community/custom
Parameters600B MoE (27B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score68License
MIT
OSI-approved
Parameters560B MoE (25B active)
FP8, Q4
Context262KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score67.9License
GLM-5.3 License
Community/custom
Parameters753B MoE (40B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score65.6License
MIT
OSI-approved
Parameters552B MoE (16B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score64.5License
MIT
OSI-approved
Parameters309B MoE (15B active)
FP8, Q4
Context1MReference hardware4× NVIDIA H100 (80GB) · 320GB total
- Score63.8License
Apache-2.0
OSI-approved
Parameters280B MoE (16B active)
FP8, Q4
Context512KReference hardware4× NVIDIA H100 (80GB) · 320GB total
- Score63.2License
MIT
OSI-approved
Parameters1600B MoE (49B active)
Q4, Q2
Context1MReference hardware8× NVIDIA H200 (141GB) · 1128GB total
- Score62.8License
Qwen Community 1.0
Community/custom
Parameters180B MoE (6B active)
FP8, Q4
Context262KReference hardware2× NVIDIA H100 (80GB) · 160GB total
- Score62.7License
Muse Model License
Community/custom
Parameters816B MoE (42B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score62.3License
MIT
OSI-approved
Parameters397B MoE (17B active)
FP8, Q4
Context262KReference hardware4× NVIDIA H100 (80GB) · 320GB total
- Score61.3License
Apache-2.0
OSI-approved
Parameters770B MoE (49B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score59.6License
Apache-2.0
OSI-approved
Parameters27B dense
BF16, Q4
Context262KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score58.5License
MIT
OSI-approved
Parameters320B MoE (18B active)
FP8, Q4
Context1MReference hardware4× NVIDIA H100 (80GB) · 320GB total
- Score58.4License
Modified MIT
Community/custom
Parameters1000B MoE (32B active)
INT4, FP8, Q4
Context256KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score56.3License
MIT
OSI-approved
Parameters744B MoE (40B active)
FP8, Q4, Q2
Context203KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score54.9License
MiniMax Community License
Community/custom
Parameters428B MoE (23B active)
FP8, Q4
Context1MReference hardware4× NVIDIA H100 (80GB) · 320GB total
- Score54.2License
Apache-2.0
OSI-approved
Parameters35B MoE (3B active)
BF16, Q4
Context262KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score53.1License
Moonshot
Community/custom
Parameters120B dense
FP8, Q4, Q2
Context256KReference hardware4× NVIDIA A100 (80GB) · 320GB total
- Score52.6License
Apache-2.0
OSI-approved
Parameters975B MoE (41B active)
FP8, Q4
Context1MReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score49.5License
Modified MIT
Community/custom
Parameters1000B MoE (32B active)
INT4, FP8, Q4
Context256KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score49.4License
OpenMDW-1.1
Community/custom
Parameters120B MoE (12B active)
FP8, Q4
Context1MReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score48.6License
Apache-2.0
OSI-approved
Parameters27B dense
FP8, Q4
Context262KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score45.7License
Apache-2.0
OSI-approved
Parameters25.2B MoE (3.8B active)
BF16, Q4
Context256KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score45.2License
OpenMDW-1.1
Community/custom
Parameters74B MoE (9B active)
FP8, Q4, INT4
Context256KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score41.2License
Apache-2.0
OSI-approved
Parameters29.6B dense
BF16, Q4
Context131KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score41.1License
MIT
OSI-approved
Parameters671B MoE (37B active)
FP8, Q4
Context128KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score39.9License
Apache-2.0
OSI-approved
Parameters31B dense
BF16, Q4
Context256KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score37.9License
LFM Open License v1.0
Community/custom
Parameters24B MoE (2.45B active)
BF16, Q4
Context131KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score37.3License
Apache-2.0
OSI-approved
Parameters117B MoE (5.1B active)
INT4
Context128KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score34.7License
Apache-2.0
OSI-approved
Parameters119B MoE (6.5B active)
FP8, Q4
Context256KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score34.5License
OpenMDW-1.1
Community/custom
Parameters398B MoE (13B active)
FP8, Q4
Context512KReference hardware4× NVIDIA H100 (80GB) · 320GB total
SourceNot linked - Score33.6License
Modified MIT
Community/custom
Parameters128B dense
FP8, Q4
Context256KReference hardware2× NVIDIA H100 (80GB) · 160GB total
- Score33.1License
Apache-2.0
OSI-approved
Parameters8B dense
BF16
Context128KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score31.9License
MIT
OSI-approved
Parameters671B MoE (37B active)
FP8, Q4
Context128KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score31.6License
Apache-2.0
OSI-approved
Parameters106B MoE (10.3B active)
FP8, Q4
Context128KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score28.7License
Qwen Research
Community/custom
Parameters72B dense
FP8, Q4
Context128KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score28.6License
Apache-2.0
OSI-approved
Parameters32B MoE (2.4B active)
BF16, Q4
Context64KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score28.2License
Llama Community
Community/custom
Parameters109B MoE (17B active)
FP8, Q4
Context10MReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score21.6License
Llama Community
Community/custom
Parameters400B MoE (17B active)
FP8, Q4
Context1MReference hardware2× NVIDIA A100 (80GB) · 160GB total
This table is the sourced deployment subset, not the complete performance ranking. A missing row means the deployment catalog is incomplete, not that the model cannot be self-hosted.
Questions
What is the best open source LLM right now?
The #1 row of the open-weight table above is the current answer, and the table recomputes from current ranking data on every build. The order can move when new evidence lands, so check the date on the ranking and the evidence label beside the leader before treating it as settled.
Are open source LLMs as good as GPT or Claude?
They can match proprietary models on individual tasks. The overall gap moves with every release, so compare the first row of the ranking above with the top of the overall leaderboard. Deployment control, privacy, and serving economics can still make an open model the better operational choice.
What is the best open source LLM for coding?
Use the open-weight rows on the live coding leaderboard. Coding order differs from overall order, and a single HumanEval or LiveCodeBench result should not be treated as a complete coding verdict.
Can I run these models locally?
The smaller open-weight rows run on a single consumer GPU with 4-bit quantization; frontier-size models need multi-GPU rigs or high-memory Apple Silicon. The local LLM guide breaks the rankings down by VRAM tier, and the Ollama guide includes pull commands and size estimates per model.
Which labs make the best open-weight models?
The current open-weight top tier comes almost entirely from Chinese labs — DeepSeek, Zhipu (GLM), Moonshot (Kimi), Alibaba (Qwen), and MiniMax — with Meta and Mistral behind on the live ranking. Family strengths differ by task, so check the per-category leaderboards and the Chinese model rankings rather than picking by lab reputation alone.
Explore More
Last updated: October 17, 2026
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.