Mistral Releases Large 4 'Le Chonk' as Open Weights, Strong at Coding but Behind China Overall
Mistral's 1 trillion parameter Large 4 is now downloadable with a modified MIT licence and cheap API pricing, though top Chinese and closed models still lead overall.
Mistral has released the weights of Mistral Large 4, nicknamed “Le Chonk”, its biggest model so far. “Open weights” means anyone can download the finished model and run or adapt it themselves, although the training data and recipe stay private. Mistral says it was trained entirely in Europe, and calls it a two year bet that is now paying off.
The numbers: about 1 trillion parameters in total (think of parameters as the model’s adjustable knobs), with roughly 52 billion in use at any moment, because it is a mixture of experts design that only wakes part of itself for each step. It takes text and images in and gives text and code back. The standard version needs eight Nvidia B300 GPUs, so this is a data centre model, not a laptop one. On the API it costs $0.70 per million input tokens and $2.10 per million output tokens (a token is a small chunk of text, roughly three quarters of a word).
On coding, Mistral reports 41% on Terminal-Bench 4.0, up from 28% in the preview, and 48% on the Artificial Analysis Coding Agent Index, about level with GLM 5.3 while using roughly 47% fewer tokens. In blind tests run by Surge, testers usually preferred its code over GLM 5.3, Kimi K3 and DeepSeek, but it beat Claude Opus 5.5 in only 22% of matchups. On the overall Intelligence Index it scores 26, behind DeepSeek V4.1 Flash and GLM 5.3 at 32. Some of these results are self reported, and a full technical report is still “coming soon”.
What’s behind it: The open model race is now largely a US and China contest, and Europe wants a seat. Mistral’s pitch is less “best in the world” and more “good, cheap, and ours to inspect”. The licence is a modified MIT: free to use, change and deploy commercially, but a company selling Large 4 as a hosted service needs a separate deal. There is also a safety note. Mistral says it refuses 96.6% of malicious cyber requests, yet admits it saw signs of possible uplift for real attacks and plans to tighten guardrails. A penetration testing variant goes only to vetted partners.
What this means for you: For most people, nothing changes today, since you cannot run a model this size at home. If you build software, it is a cheap, flexible option to test against your current coding assistant, especially if you care about where your data goes or want European hosting. If you are curious about open models, watch the smaller releases that tend to follow big ones. And keep expectations grounded: it is the best of its kind in some areas, not the best overall.
Sources
OpenAI Doubles Usage Limits on Its Pro Plans and Reverses a Planned Cut
OpenAI has doubled the Work and Codex quotas on its Pro 100 and Pro 500 plans and dropped a planned reduction on Pro 200, as the race with Anthropic heats up.
