Elite Open Source AI Model¶
The definitive ranking of the world's most capable open-source / open-weight AI models — the frontier leaderboard built from official releases and independent benchmarks.
What this is¶
A continuously updated reference for the open-weight frontier: models you can self-host, fine-tune, and build on. The ranking covers the strongest open-weight releases from Chinese labs (DeepSeek, Moonshot/Kimi, Alibaba/Qwen, Zhipu/GLM, MiniMax) as well as Mistral AI, Meta, and OpenAI.
Every model page includes:
- Quick facts — rank, developer, country, release date, license (with OSI status, commercial use, self-hosting)
- Architecture — MoE vs dense, total & active parameters, context window
- Benchmarks — Artificial Analysis Intelligence Index, SWE-bench, GPQA Diamond, Terminal-Bench and more
- API pricing — input/output cost per million tokens in USD
- Best for and release notes
The current leader¶
Rank #1 open-weight model
Kimi K3 (Moonshot AI) — the first open-weight model in the ~3T parameter class, and per August 2026 the #1 open-weight model on the Artificial Analysis Intelligence Index.
Explore¶
- Leaderboard — full comparative table of all ranked models
- All models — browse the complete catalog
- Methodology — how the ranking is built, sources, and caveats
- Raw data — download the underlying JSON dataset
Quick start¶
If you just want a practical recommendation:
- Best overall — Kimi K3 (rank #1) or GLM-5.3 (rank #2)
- Best for low-cost coding — DeepSeek V4-Pro (MIT, rank #3)
- Best for cheap, high-volume agentic workloads — DeepSeek V4-Flash (rank #4)
- Best for self-hosting on a single GPU — Qwen3.8-27B (Apache 2.0, rank #6) or gpt-oss-120b (rank #13)
- Best for European data compliance — Mistral Large 3 (Apache 2.0, rank #10)
Open-weight ≠ open-source
"Open-weight" means the model weights are publicly released, but the license may not be OSI-approved (e.g. Kimi's Modified MIT, Llama's Community License). Always check the license before commercial use.