Mistral releases the open weights of Mistral Large 4, its 1-trillion-parameter AI model

French AI company Mistral AI on Thursday 29 October released the weights of Mistral Large 4, its flagship model, which anyone can now download from Hugging Face and run on their own computers. The model, nicknamed “Le Chonk”, had been in public preview since earlier in October.

Mistral said three more weeks of reinforcement learning since the preview had improved the model’s coding and defensive cybersecurity skills and made its reasoning more efficient. It described Large 4 as “the best open-weight model for coding”. These are Mistral’s own claims. The company said the model’s score on Terminal-Bench 4.0, a test of coding in a computer terminal, rose from 28% at preview to 41%.

Mistral says Large 4 has 1 trillion parameters, of which 52 billion are used for each task, and accepts text and images. The Next Web reported that it was trained on 3,800 Nvidia Grace Blackwell chips in Europe. Users can now choose low, medium or high reasoning modes.

On cybersecurity, Mistral said the model refused 96.6% of malicious cyber requests on average in its tests. It also said it had seen “a potential for uplift to conduct real attacks” in live use since the preview and was changing its guardrails. A specialised variant for penetration testing is available only to vetted enterprise partners, it said.

The model is released under a modified MIT licence that allows free use, modification and commercial use; companies offering it as a hosted service must buy a separate licence. Mistral listed API prices of $0.70 per million input tokens and $2.10 per million output tokens. The model is also available through Microsoft Foundry.

“Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion,” chief executive Arthur Mensch said, according to The Next Web.