Loading benchmark data…
EyesInAI·Loading live benchmark data
Open-weight models you can run yourself, and the tools to run them — curated picks plus every free open model we benchmark.
Several of the strongest models below are from Chinese labs. The data risk is decided by how you run them — self-hosted weights send nothing to China; a China-hosted API puts your data under PRC law. The three deployment modes by risk, plus a checklist to diminish each.
Where open models stand right now — ranked by benchmark, mapped to the hardware you need, and matched to the job. Chinese labs hold the entire top tier; the best open model sits 4 points behind the top proprietary model on composite.
Tap any model for its benchmark breakdown, strengths, and self-host footprint.
Our tests = EyesInAI’s own live benchmark — pass-rate across our suite for the best-performing vendor of each model, updated as the bench re-runs. ⚠ flags where our measured result lags the reported hype. The composite Score and the reported per-benchmark figures are third-party (BenchLM.ai composite, May 2026, cross-checked against published reviews); where sources differ we show a range or “~”, and figures we couldn’t corroborate were dropped or flagged. Best proprietary for reference: Gemini 3.1 Pro 91 · GPT-5.4 88 · Claude Opus 4.6 86.
Dead-simple local model runner. github.com/ollama/ollama — pull and run gemma, llama, qwen, etc. in one command.
Open repoHigh-throughput local inference server. github.com/vllm-project/vllm — the standard for serving open models fast on your own GPU.
Open repoOpen-weight models available free through our providers — click any for its scorecard.