Mistral Large 4 open weights broaden defensive AI deployment choices

Mistral has released Large 4 weights after its public preview, with new coding and cyber benchmarks. Security teams should test the model in controlled workflows before trusting its defensive claims.
Mistral has released the weights for Mistral Large 4, its one trillion parameter multimodal model, following a public preview earlier in October. The company says additional training improved coding, defensive cyber tasks and reasoning efficiency. For security teams, the practical change is the ability to evaluate and host a frontier scale model under their own controls rather than relying solely on a hosted endpoint.
What the release adds
Mistral reports a 48% score on the Artificial Analysis Coding Agent Index, tied with the leading open weight model in its comparison while using 47% fewer tokens per task. It also reports 41% on Terminal-Bench 4.0, a 13 percentage point increase from the preview. These are vendor reported results, and the benchmark mix may not resemble an organisation's codebase or incident workflow.
On the Artificial Analysis Cyber Index, Mistral reports 51% for tasks spanning vulnerability discovery, reproduction and patching. It says the model refused 96.6% of malicious cyber requests in its evaluation. That refusal figure is relevant to risk reviews, but it does not establish that every deployment configuration or downstream agent is safe.
We release the weights of Le Chonk.
Deployment choices have conditions
The weights are available for self hosting and customisation, while Mistral also offers an API and managed routes. Mistral describes its Ecosystem Open-Weight licence as a modified MIT licence: commercial use is allowed, but businesses offering the model as a service need a separate commercial agreement. Procurement teams should review the licence against their actual delivery model before rollout.
Three reasoning modes let teams trade response cost against precision. Low is aimed at simple tasks, medium at routine work and high at harder cases. Mistral also highlights a specialised cyber variant for vetted enterprise partners, which is distinct from general access to the released weights.
What defenders should test
The clearest first use is an isolated vulnerability triage pilot. Give the model a bounded repository, known issues and a test harness; measure whether suggested patches compile, preserve behaviour and remove the flaw. Keep exploit reproduction within owned or explicitly authorised systems. Review refusal behaviour and tool permissions separately from coding accuracy.
Security leaders should record data residency, model provenance, access controls and patch approval in the same evaluation. Open weights make local inspection and deployment possible, but they do not replace logging, human review or independent testing. Mistral's release expands the choice set; its operating value will depend on measured performance in each organisation's own environment.
Sources
Written by the Devence Lab research team.

