French AI company Mistral AI has opened downloads of its Mistral Large 4 model while keeping a cyber-focused version inside a vetted partner program, giving organizations local access as the company adjusts safeguards against misuse.
The weights became available Oct. 29, following a public preview announced Oct. 6. Organizations can download them and run the model on their own systems. Mistral describes Mistral Large 4 as a multimodal model with 1 trillion parameters. Its model card lists 49 billion active parameters per token, or 52 billion when embeddings and output layers are included. Mistral says a cluster of 3,800 NVIDIA Grace Blackwell GPUs at its European data centers powered the training run.
The specialized version is being previewed with vetted enterprise design partners for advanced cybersecurity work. Mistral says testing is limited to systems partners own or are authorized to assess and is subject to contractual safeguards.
Mistral reports a 51% result on the Artificial Analysis Cyber Index. Mistral says the Cyber Index evaluates vulnerability discovery, validation and patching. Mistral also reports that the model refused an average 96.6% of malicious requests across JailbreakBench, StrongREJECT and AgentHarm. The refusal figure does not measure real-world attack outcomes.
The company says monitoring of live traffic identified potential for the model to assist real attacks, even as it observed refusals of black-box attacks against live targets. Mistral is revising safeguards and plans further updates in the coming months. It also says a technical report is forthcoming, without giving a publication date.
Mistral says local access lets security teams set their own policies for defensive investigations, including work a hosted service might refuse. The weights are distributed under the Mistral Ecosystem Open-Weight license, which the company describes as a modified MIT license. Businesses operating a Model-as-a-Service offering need a separate commercial license.
