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

From public preview to open weights

Mistral Large 4 first appeared earlier this month in public preview as the company's flagship model. Mistral says the response exceeded its expectations and that it used developer and customer feedback while continuing to scale reinforcement learning. The result, according to the company, is a set of large improvements in coding and defensive cybersecurity, with gains that show "no signs of saturation".

The model was trained from scratch on European infrastructure, and Mistral stresses its clear lineage from inception to release. The company presents this as a trusted base for what it calls sovereign AI, where organizations own, customize and control their models. The new version also brings better reasoning efficiency, which Mistral says gives developers more performance at lower inference costs.

Coding results and reasoning modes

Mistral reports that three more weeks of reinforcement learning moved the model to the front of open-weight coding. It says Mistral Large 4 takes the lead on FrontierCode and Terminal-Bench. On Cognition's private FrontierCode 2.0 benchmark, which judges end-to-end code quality and whether a repository maintainer would merge an agent's change, it scores 19.7%.

• On the Artificial Analysis Coding Agent Index, the model scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average, according to Mistral.
• On Terminal-Bench 4.0, it improved by 13 percentage points to 41% compared with the public preview, while cutting output tokens by 30%.
• In Surge Code's blind pairwise evaluations, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked ahead among the models evaluated.

Developers can now choose between three reasoning modes. Low is meant for simple tasks and replaces the earlier 'None' mode. Medium suits everyday tasks. High is the default and targets the hardest use cases that need high precision and thorough reasoning.

Cybersecurity and safety claims

Mistral says it worked with cybersecurity vendors since the preview to test the model on vulnerability reproduction and patching. It scores 51% on the Artificial Analysis Cyber Index, which measures how well a model finds vulnerabilities in a codebase, reproduces and validates them, and patches them without breaking existing functionality. The company says the model handles legitimate defensive tasks without unnecessary refusals.

On the safety side, Mistral says built-in guardrails separate defensive from malicious, offensive requests. It reports an average refusal rate of 96.6% on malicious cyber requests, measured on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm. The company calls this the highest it measures among open models.

Industry reactions and what it means

Mistral quotes several security organizations. Brad Medairy of Booz Allen Hamilton said testing showed the model matched vulnerability detection recall rates of other frontier cyber models "with a compelling cost curve". Elastic's James Spiteri said Mistral's open models are supported on Elastic Security today. ENISA's Laura Heuvinck said the agency has access to the model and is testing it to identify security vulnerabilities, audit code and analyse defensive logs. Cisco, Cloudflare and CrowdStrike representatives also commented, stressing transparency, control and self-hosting.

For organizations that need to host and adapt AI on their own infrastructure, the release of the weights makes Mistral Large 4 an option for coding agents and defensive security work. All benchmark figures come from Mistral's own announcement and cited evaluators.