Mistral Large 4 passes in open weight: three more weeks of training and a 13-point jump on Terminal-Bench

The model of a trillion parameters of Mistral is now downloadable under modified MIT license, with net progress in code and defensive cybersecurity since its preversion.

Three weeks after its public preversion, Mistral AI publishes the weights of Mistral Large 4, nicknamed "the Chonk". This multimodal expert mix model has a trillion parameters (about 52 billion active according to the name of the Hugging Face repository), brings together instruction and reasoning modes, and covers more than 160 languages. Mistral recalls that it has been driven from zero on European infrastructures.

What has changed since preversion

The company devoted an additional three weeks of reinforcement learning to code and defensive cybersecurity. On Terminal-Bench 4.0, the score rose from 28% to 41% while generating 30% less chips; Resistance to attacks against agents (B3) increased from 93.3% to 97.1%, and the refusal of malicious cyber queries reached 96.6%. Three levels of reasoning (low, medium, high by default) are proposed. Mistral presents it as the best open model for code, while acknowledging that it has observed a potential gain for real attackers: an offensive security-oriented variant remains reserved for verified partners, such as Orange Cyberdefense.

The licence, an amended TIM, authorizes use, modification and resale, except for those undertakings that would offer it in hosted service, subject to a commercial licence. Via the API, it costs $0.70 per million tokens in entry and $2.10 in output; It is also available from Nvidia, OpenRouter and Microsoft Foundry.