Mistral releases Large 4 weights after additional training

On 29 October 2026, Mistral released the weights of Mistral Large 4, also known as Le Chonk. The move turns the model from an API preview into an open-weight release that organizations can download, inspect and run on their own infrastructure. It is the material update to Mistral’s original 6 October announcement: the earlier version offered a preview and promised weights by the end of the month, while this release provides an improved checkpoint after roughly three additional weeks of reinforcement learning and customer testing.

Mistral describes Large 4 as a natively multimodal model with about one trillion parameters and 52 billion active parameters. The company says the released version is trained for coding, agentic workflows and defensive cybersecurity, and now supports three reasoning modes: low, medium and high. That gives developers a way to trade speed and output cost against more deliberate reasoning, although self-hosting still requires substantial hardware rather than an ordinary laptop.

The practical change is control. An API preview lets a provider manage infrastructure and access terms; published weights allow a company, public institution or researcher to deploy the model inside its own environment and adapt it to local requirements. That matters for regulated work and security operations, but open weights are not the same as unrestricted use. Hardware, inference engineering, the license and safeguards still determine what a deployment can actually do.

The update also includes fresh performance claims. Mistral reports 19.7% on Cognition’s private FrontierCode 2.0 evaluation, 48% on the Artificial Analysis Coding Agent Index with 47% fewer tokens per task than the comparison point, and 41% on Terminal-Bench 4.0, up from 28.3% in the preview. It reports 51% on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests. These are company-reported results; benchmark setups, model versions and independent reproduction matter before treating them as general rankings.

For Europe’s AI ecosystem, the release is significant because it combines frontier-scale capability with a self-hostable distribution model. For developers, the real test now moves from preview access to deployment: whether the weights, license, hardware cost and safeguards make Large 4 useful on concrete workloads. The release is consequently more than a larger model announcement; it is a change in who can control the model after downloading it.