Score66.8
License
MIT
OSI-approved
Parameters
744B MoE (40B active)
FP8, Q4, Q2
Context203K
Reference hardware
8× NVIDIA H100 (80GB) · 640GB total
BenchLM recommendation
Open-Source LLM Leaderboard 2026
Data verified
The best open-source LLM by current open-weight benchmark score is Qwen3.8 Max. It leads the September 2026 open-weight ranking at 78.7, ahead of Hy4 preview (78.3) and Qwen3.8-27B (72). The license directory below separates OSI-approved licenses from community terms.
Citable stat 106 open-weight models are ranked as of September 2, 2026; Qwen3.8 Max leads at 78.7/100.
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.
About this ranking
Last verified: September 2, 2026
This page is the canonical open-weight ranking. It uses the same public BenchAlign v5 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 overall lane filtered to open-weight models. Scores measure capability; evidence labels and score intervals show how much confidence to place in close comparisons.
The open-weight slice starts with Qwen3.8 Max, followed by Hy4 preview and Qwen3.8-27B. All rows use the public BenchAlign v5 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 overall contract filtered to open-weight models. For detailed model profiles, click any model name below. 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 12 of 12 deployment-documented rows
Deployment catalog checked 2026-06-12.
- Score60.2License
Apache-2.0
OSI-approved
Parameters31B dense
BF16, Q4
Context256KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score59.3License
Modified MIT
Community/custom
Parameters1000B MoE (32B active)
INT4, FP8, Q4
Context256KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score59.2License
Modified MIT
Community/custom
Parameters1000B MoE (32B active)
INT4, FP8, Q4
Context256KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score58.9License
Moonshot
Community/custom
Parameters120B dense
FP8, Q4, Q2
Context256KReference hardware4× NVIDIA A100 (80GB) · 320GB total
- Score53.7License
Apache-2.0
OSI-approved
Parameters27B dense
FP8, Q4
Context262KReference hardware1× NVIDIA RTX 4090 (24GB) · 24GB total
- Score53.1License
Qwen Research
Community/custom
Parameters72B dense
FP8, Q4
Context128KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score51.1License
MIT
OSI-approved
Parameters671B MoE (37B active)
FP8, Q4
Context128KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score46.5License
Apache-2.0
OSI-approved
Parameters119B MoE (22B active)
FP8, Q4
Context256KReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score44.5License
MIT
OSI-approved
Parameters671B MoE (37B active)
FP8, Q4
Context128KReference hardware8× NVIDIA H100 (80GB) · 640GB total
- Score39.7License
Llama Community
Community/custom
Parameters109B MoE (17B active)
FP8, Q4
Context10MReference hardware1× NVIDIA H100 (80GB) · 80GB total
- Score22.9License
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.
What changed
Qwen3.8 Max leads the live open-weight ranking at 78.7 with Supported evidence.
Hy4 preview ranks #2 at 78.3 with Estimated evidence.
Qwen3.8-27B ranks #3 at 72 with Supported evidence.
How to choose
Best open-weight model overall?Start with the live #1, then check its evidence label and deployment constraintsCoding-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
Full Rankings (106 models)
4
dots3-note Preview
Dots Studio·Open Weight·512K
68.7
BenchAlign v5
Estimated
90% interval 58.80–78.54
11
Inkling
Thinking Machines Lab·Open Weight·1M
66.7
BenchAlign v5
Supported
90% interval 59.06–74.26
13
Inkling-Small
Thinking Machines Lab·Open Weight·1M
63.6
BenchAlign v5
Supported
90% interval 57.95–69.23
17
Agents-A1-F16-GGUF
InternScience·Open Weight·262K
61.2
BenchAlign v5
Estimated
90% interval 51.35–71.09
18
Agents-A1-FP8
InternScience·Open Weight·262K
61.2
BenchAlign v5
Estimated
90% interval 51.35–71.09
19
Agents-A1-Q4_K_M-GGUF
InternScience·Open Weight·262K
61.2
BenchAlign v5
Estimated
90% interval 51.35–71.09
20
Agents-A1-Q8_0-GGUF
InternScience·Open Weight·262K
61.2
BenchAlign v5
Estimated
90% interval 51.35–71.09
21
Qwen3.8-Flash-Next
Alibaba·Open Weight·262K
61.1
BenchAlign v5
Estimated
90% interval 49.54–72.57
24
Qwen3.5 397B (Reasoning)
Alibaba·Open Weight·128K
60.5
BenchAlign v5
Estimated
90% interval 48.99–72.02
28
Kimi K2.7 Code
Moonshot AI·Open Weight·256K
59.3
BenchAlign v5
Estimated
90% interval 48.33–70.24
30
DeepSeek V3.2 (Thinking)
DeepSeek·Open Weight·128K
59.1
BenchAlign v5
Estimated
90% interval 47.61–70.63
31
Qwen3 235B 2507 (Reasoning)
Alibaba·Open Weight·128K
59
BenchAlign v5
Estimated
90% interval 47.49–70.52
36
Trinity-Large-Preview
Arcee AI·Open Weight·512K
56.9
BenchAlign v5
Estimated
90% interval 45.38–68.41
44
Ling 3.0 Flash
InclusionAI·Open Weight·262K
53.8
BenchAlign v5
Estimated
90% interval 42.27–65.30
47
Ornith-1.5-35B-A3B
Ornith AI·Open Weight·262K
53.4
BenchAlign v5
Estimated
90% interval 43.50–63.24
50
DeepSeek V3.1 (Reasoning)
DeepSeek·Open Weight·128K
52.9
BenchAlign v5
Supported
90% interval 33.10–72.69
52
Nemotron 3 Super 120B A12B
NVIDIA·Open Weight·256K
51.9
BenchAlign v5
Estimated
90% interval 40.34–63.37
55
DeepSeek Coder 2.0
DeepSeek·Open Weight·128K
51.2
BenchAlign v5
Estimated
90% interval 39.64–62.67
62
Ministral 3 14B (Reasoning)
Mistral·Open Weight·128K
50.2
BenchAlign v5
Estimated
90% interval 38.69–61.72
63
Mixtral 8x22B Instruct v0.1
Mistral·Open Weight·64K
49.4
BenchAlign v5
Estimated
90% interval 37.90–60.92
65
Trinity-Large-Thinking
Arcee AI·Open Weight·512K
48.1
BenchAlign v5
Supported
90% interval 30.81–65.33
73
Nemotron 3 Nano Omni 30B A3B
NVIDIA·Open Weight·256K
44.7
BenchAlign v5
Estimated
90% interval 33.16–56.19
75
Ling 2.6 Flash
InclusionAI·Open Weight·262K
44.4
BenchAlign v5
Estimated
90% interval 32.85–55.88
80
DeepSeek R1 Distill Qwen 32B
DeepSeek·Open Weight·128K
43.1
BenchAlign v5
Estimated
90% interval 31.57–54.60
87
Ministral 3 8B (Reasoning)
Mistral·Open Weight·128K
41.1
BenchAlign v5
Estimated
90% interval 29.62–52.65
88
Exaone 4.0 32B
LG AI Research·Open Weight·128K
41
BenchAlign v5
Estimated
90% interval 29.52–52.55
92
Ministral 3 3B (Reasoning)
Mistral·Open Weight·128K
40.3
BenchAlign v5
Estimated
90% interval 28.75–51.77
94
Exaone 4.0 1.2B
LG AI Research·Open Weight·128K
39.7
BenchAlign v5
Estimated
90% interval 28.21–51.24
98
Qwen2.5 Coder 32B Instruct
Alibaba·Open Weight·128K
34.4
BenchAlign v5
Supported
90% interval 17.63–51.09
101
Nemotron 3.5 Lightning 30B A3B NVFP4
NVIDIA·Open Weight·1M
26.6
BenchAlign v5
Estimated
90% interval 16.73–36.46
Key Takeaways
The top model is Qwen3.8 Max by Alibaba with a BenchAlign v5 score of 78.7 and Supported evidence.
The best open-weight model is Qwen3.8 Max at position #1.
106 models are included in this ranking.
Score in Context
What these scores mean
Open-weight models are ranked by the same public BenchAlign v5 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.
Best Open Source LLMs FAQ
What is the best open source LLM right now?
The live table above is the ranking owner and recomputes from the public BenchAlign v5 artifact. MiniMax M3 leads the July 14 snapshot with Supported evidence; use the table rather than a copied winner sentence when the data changes.
Are open source LLMs as good as GPT or Claude?
They can match proprietary models on individual tasks, but the July 14 unified overall ranking still shows a double-digit gap between the open-weight and proprietary leaders. 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: September 2, 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 2,000+ readers.
One email each week. Unsubscribe anytime.