Mistral releases open weights for Mistral Large 4, a 1-trillion-parameter model

Mistral has released the weights of Mistral Large 4, nicknamed "Le Chonk," a 1-trillion-parameter, natively multimodal model. The model was first introduced in public preview earlier this month. Businesses and governments can now download and deploy it themselves. Weights are on Hugging Face, and the model is also available through Mistral's API.

Mistral says the released version is improved over the preview, crediting three more weeks of reinforcement learning. It reports gains in coding, defensive cybersecurity and reasoning efficiency. The model now offers three reasoning modes, letting developers trade cost against performance, with the top mode aimed at the hardest use cases that need high precision.

On coding, Mistral says the model scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task. It reports 41% on Terminal-Bench 4.0, up 13 points from the preview, and 19.7% on Cognition's private FrontierCode 2.0, which it says is the lead among open-weight models. In Surge Code's blind pairwise evaluations, Mistral says the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5.

On security, Mistral reports 51% on the Artificial Analysis Cyber Index, which covers finding, reproducing and patching vulnerabilities. It says the model handles legitimate defensive tasks without unnecessary refusals while refusing malicious cyber requests 96.6% of the time on average. Booz Allen Hamilton and Cisco's AI threat research group gave supportive statements. The benchmark figures are Mistral's own claims.

Mistral is arguing that open weights matter most where control matters most. Pierre Stock, Mistral's vice president of science, told TechCrunch that open weights give organizations access to frontier capability that cannot be cut off. He also said organizations can run the model on-premises. For governments and security teams with data-residency requirements, that is a practical difference from API-only models.

The comparison with GLM-5.3 is notable because that Chinese model was previously considered among the best open-weight options, particularly for cyber work. A similar score at 47% fewer tokens per task, if it holds up in independent use, would make agent workloads cheaper to run. Mistral also stresses its European training infrastructure and clear lineage as selling points for sovereign AI.

Stock framed wide availability as defensive: security tools that find bugs faster should reach everyone quickly, he said, because malicious actors may use AI to attack. The tradeoff is that open weights cannot be recalled once released, so the refusal-rate claim is one to test independently.

Independent evaluations will show whether the coding and cyber results hold outside Mistral's reported figures. FrontierCode 2.0 is a private benchmark, which limits outside checking.

Also worth watching is how enterprises and public-sector buyers adopt self-hosted deployments, and how safety researchers assess the model's behaviour once the weights are widely available.