Mistral releases Large 4’s open weights, claiming coding gains with fewer tokens

Mistral AI released the weights of an improved Mistral Large 4 on October 29, moving its flagship model beyond the public preview introduced earlier this month. The release lets organizations download and run the model themselves, while Mistral claims stronger coding performance with fewer tokens—and acknowledges that its cybersecurity safeguards still need adjustment.

Nicknamed Le Chonk, Large 4 is a 1-trillion-parameter model that Mistral describes as natively multimodal. The company says it trained the model from scratch on European infrastructure. Its pitch centers on ownership and control: organizations can customize the model and deploy it on their own infrastructure, rather than rely solely on a hosted service.

The weights are available through Hugging Face, with self-hosted deployment supported through vLLM or SGLang, software used to run models. Hosted access is also available through Mistral Studio, Mistral’s API and OpenRouter. The announcement links to a model card and license, but does not spell out the license terms, context window or token prices.

A coding upgrade measured in answers and tokens

Mistral credits three additional weeks of reinforcement learning—training that rewards successful attempts—for the coding improvements. On the Artificial Analysis Coding Agent Index, it reports a 48% score, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. Tokens are the units of text a model processes or generates; fewer can reduce the work needed to complete a task.

Terminal-based coding shows a sharper change from the preview. Mistral reports a 13-percentage-point gain on Terminal-Bench 4.0 while cutting output tokens by 30%. These are results presented by Mistral, not a guarantee of equivalent savings on a customer’s workload.

A separate test targets whether code is worth accepting, rather than merely whether an agent finishes. Mistral reports 19.7% on Cognition’s private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge the change. The company also cites Surge Code’s blind comparisons, where Large 4 ranked behind only Claude Opus 5.5 among the models evaluated.

Developers now have three reasoning settings to balance performance and cost. Mistral recommends:

• Low for simple tasks, replacing the previous None setting.
• Medium for everyday work.
• High, the default, for the hardest tasks requiring precision and thorough reasoning.

Defensive capability, with safeguards still changing

For cybersecurity, Mistral reports 51% on the Artificial Analysis Cyber Index. That test covers finding vulnerabilities, reproducing and validating them, and patching code without breaking existing functionality. The company says Large 4 handles legitimate defensive tasks without unnecessary refusals, while its guardrails distinguish those requests from malicious activity.

Mistral reports a 96.6% average refusal rate on malicious cyber requests across cyber-related prompts from three safety tests. But it also says monitoring live traffic revealed potential for the model to help conduct real attacks, even though it observed refusals of black-box attacks against live targets. Mistral says it is modifying its guardrails and will keep updating them in the coming months.

A specialized cybersecurity variant has tighter access conditions. Enterprise design partners, including Orange Cyberdefense, are receiving preview access for penetration testing, vulnerability assessment and red-teaming. Mistral says access requires vetting and contractual safeguards, with testing limited to owned or explicitly authorized systems in controlled environments.