Mistral Large 4 Brings a Trillion-Parameter Open-Weight Bet to Europe

Mistral Large 4 is the French lab’s biggest model yet. Mistral unveiled it on 6 October. The model has about one trillion parameters in total, but only 49 billion are active for each token.

That design is a mixture-of-experts system. Only part of the network runs at a time. It cuts the cost of answering a prompt while keeping a large pool of knowledge. SiliconANGLE reports that the model supports more than 160 languages.

What Mistral Large 4 offers developers

A public preview is available now through Mistral’s cloud platform. The company says the model weights will follow later in October. Open weights would let outside teams run and adapt the model themselves. That is the main promise to European users who want alternatives to US providers.

Mistral trained the model on nearly 4,000 Nvidia Grace Blackwell chips. It also built a software stack that runs tens of thousands of rollouts in parallel. The company says this generated about 33 billion tokens a day and let training continue without pauses.

Benchmark claims need caution. Mistral says the model scored 82 percent on the AA Cyber Index for patching open-source vulnerabilities. It also says the model beat GPT-6 Astra by one percentage point on a computer vision test. It claims leads over Qwen3.8 Max and DeepSeek V4 Pro on coding. These are company figures. Independent tests will matter more.

The cyber angle is notable. Anthropic expanded its own verification programme for security defenders this week. Both moves show labs treating software security as a core use case.

Mistral also says larger and more capable versions will arrive in the coming months. It plans a series of models tuned for specific jobs. Pricing and licence terms were not specified in the coverage we reviewed. Developers should wait for the final release notes before planning production use.

The competition is crowded. Nvidia-backed rivals are also chasing open-weight leadership, as we reported in our Reflection AI report.

Fact basis: SiliconANGLE, 6 October 2026 and AI Weekly, 7 October 2026. Benchmark figures are vendor-reported.