Mistral AI has released the weights of Mistral Large 4, the 1-trillion-parameter multimodal model it put into public preview earlier this month. The company says three more weeks of reinforcement learning pushed it to the front of open-weight coding benchmarks and into the top tier of models used for defensive security work.
The number that matters is not the parameter count. A trillion-parameter model published openly can be downloaded, fine-tuned and served inside an organisation's own infrastructure — the difference between renting access to a foreign API and owning the capability. Mistral, a French lab, is explicit that this is the pitch: the model was trained from scratch on European infrastructure with a documented lineage from inception to release, which is the argument European governments and regulated industries have been waiting for.
What the benchmarks say
Mistral reports that Mistral Large 4 now leads open-weight models on Cognition's private FrontierCode 2.0 benchmark, which tests end-to-end code quality and whether a repository maintainer would merge the agent's change, at a score of 19.7 per cent. On the Artificial Analysis Coding Agent Index it scores 48 per cent — effectively tied with GLM-5.3, the leading open-weight model, while using 47 per cent fewer tokens per task. On Terminal-Bench 4.0 it improved 13 percentage points over its preview to 41 per cent, again leading open-weight models, with 30 per cent fewer output tokens.
On the Artificial Analysis Cyber Index, which measures whether a model can find vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality, it scores 51 per cent. Mistral says the model handles those legitimate defensive tasks without excessive refusals — a claim that matters as much as the score, because a security model that refuses real work is useless to the team that needs it.
Three reasoning modes let a deployment trade quality against inference cost per request. For API customers the model is priced at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.
The licence has one seam
The weights ship under the Mistral Ecosystem Open-Weight licence, a modified MIT licence. Anyone can use, modify, fine-tune and commercialise the model under one condition: a business running a model-as-a-service offering must obtain a separate commercial licence from Mistral first. The carve-out leaves researchers, enterprises and self-hosters with an open model while stopping a competitor from reselling Mistral's own work back to it at a lower price.
Defence is the product
Mistral is not presenting this as a general-purpose release. The security framing runs through the announcement: a model that can read a codebase, reproduce a vulnerability and patch it is the thing security teams have been told they would eventually get from AI, and Mistral is claiming it first among open models. Booz Allen tested the model against an internal benchmark for source code security analysis and, according to Mistral, matched the vulnerability-detection recall of other frontier cyber models on a better cost curve. Enterprises can adapt it to proprietary data through Mistral Forge, and NVIDIA's NeMo AutoModel and TensorRT-LLM are the documented paths for customisation and high-throughput serving.
What is not settled
The benchmark figures are Mistral's own, or were run on benchmarks whose results Mistral chose to publish; independent replication is the next test, not a formality. More important is the dual-use question. The same ability to find a vulnerability in a codebase is the ability to find it for someone who intends to exploit it, and a 51 per cent score on a defensive index does not say where the refusal line falls in practice. Mistral's position — capable on defence, refusing on harm — is a design goal until someone outside the company measures it.
Until then, the release changes the floor: the most capable open-weight security model on the market is now one that any organisation can host itself, and the argument that frontier capability requires a closed API has one fewer example behind it.
Sources: Mistral — Mistral Large 4 (Le Chonk) is now open weight; Artificial Analysis.