Mistral Large 4 goes open-weight with big coding gains

Mistral AI published the weights of Mistral Large 4 on Hugging Face on Thursday, 29 October 2026. The model, nicknamed "Le Chonk", has one trillion parameters and can now be run on a company's own servers. The release comes three weeks after its 6 October debut, when it was available only through a paid API.

Weights are the settings a model learns during training. Downloading them lets users run it without sending data to Mistral. Large 4 uses a mixture-of-experts architecture with 128 specialised experts. Only four of them, plus one shared expert, are active for each token, so the model uses 49 billion parameters per token (52 billion including input and output layers). Its predecessor, Mistral Large 3, had 675 billion parameters, 41 billion of them active.

What changed since the preview

Mistral kept up reinforcement learning during those three weeks, mainly on code and cybersecurity. According to the company, its Terminal-Bench 4.0 score rose from 28% to 41%, one point ahead of China's GLM-5.3 at 40%. It also ahead of DeepSeek V4.1 Flash (31%), Kimi K3 (21%) and Muse Spark Open (6%). On Artificial Analysis's coding index it scores 48%, level with GLM-5.3, while using on average 47% fewer tokens per task.

Mistral concedes in its Hugging Face documentation that the best Chinese open models still lead on Artificial Analysis's overall index. Large 4 scores 26 points there, the best among Western open models. Search engine Ecosia replaced Mistral with Chinese models such as Qwen, Kimi and GLM in early October.

Cheaper API, tighter licence

API pricing has been cut roughly in half, to $0.70 per million input tokens and $2.10 per million output tokens, down from $1.36 and $4.18 in the preview.

The licence is less permissive than the Apache 2.0 used for Large 3. The new Mistral Ecosystem Open-Weight License, a modified MIT licence, allows commercial use, but companies that resell access to the model as a service must sign a separate agreement with Mistral.

The model remains out of reach for personal computers. Even the compressed version needs four Nvidia B200 or B300 data-center GPUs, and the full version needs eight.