Open Source LLM Leaderboard

Which open-source models appear on this leaderboard?

The leaderboard tracks open-weight or publicly available models — including Llama, Qwen, DeepSeek, Mistral, GLM, and other releases whose weights or code are available under tracked licenses. It may include permissive, non-commercial, or otherwise restricted licenses; closed-weight API-only models such as GPT or Claude are excluded here.

Why do scores for the same model differ across benchmarks?

Each benchmark measures a different capability — reasoning (HLE, ARC-AGI-2), math (AIME, FrontierMath), coding (SWE-bench Verified), agent use (τ²-Bench), and so on. A model tuned for one capability may perform very differently on another, which is exactly why we surface per-benchmark scores rather than a single number.

How often is the leaderboard updated?

Data is revalidated every 5 minutes, and new models or evaluation results are added as soon as they are published. The "Updated on" indicator at the top of the page reflects the most recent data refresh.

How should I read the composite ranking?

The composite view aggregates a model's standing across multiple core benchmarks. It is a useful first filter, but for production decisions you should drill into the specific benchmark closest to your workload — for example, SWE-bench Verified for coding agents, or τ²-Bench for tool-use scenarios.

Can I run these open-source models locally?

Most listed models publish weights on Hugging Face or GitHub and can be served via vLLM, Ollama, llama.cpp, or similar runtimes. Hardware requirements scale with parameter count — a 7B model fits on a single consumer GPU, while 65B+ models typically need multi-GPU or quantized deployment.