Mistral Large 4 is open-weight: self-host a 1T near-frontier coder
Its near-MIT license lets you fine-tune and ship commercially — only API resale needs Mistral's OK. The specs, the coding scores, and how to run it today.
Copy markdown
Le Chonk: a 1T MoE you download and run
Mistral Large 4 is a roughly 1-trillion-parameter Mixture-of-Experts (about 52B active, per the 'A52B' repo), natively multimodal, with Low/Medium/High reasoning modes across 160+ languages. The weights are on Hugging Face — serve them with vLLM or SGLang, or pull the NVIDIA NIM / TensorRT-LLM build.
The license: commercialize freely, just don't resell raw inference
It ships under the Mistral Ecosystem Open-Weight license — a modified MIT that lets you use, modify, fine-tune and commercialize it inside your own product. The one catch: running it as a Model-as-a-Service for others needs a separate commercial license from Mistral. Your internal agent or app is clear.
Coding near the frontier, at roughly half the tokens
Mistral reports 48% on the Artificial Analysis Coding Agent Index — effectively tied with GLM-5.3 while burning ~47% fewer tokens per task — plus 41% on Terminal-Bench 4.0, up 13 points from the preview. On Surge's human-preference coding eval it trailed only Claude Opus 5.5 among models tested. Vendor numbers, not independently verified.
Rather not host a 1T model? It's cents on the API
Hosted inference runs $0.70 per 1M input tokens ($0.07 cached) and $2.10 per 1M output — undercutting most closed frontier coders. It's live on OpenRouter, Mistral Studio, NVIDIA endpoints and Microsoft Foundry, so you can A/B it against your current model without touching a GPU.