Summary
Mistral has opened an API preview of Mistral Large 4, a multimodal mixture-of-experts model with one trillion total parameters and 49 billion active at a time. Its weights are scheduled for release on October 27, including FP8 and FP4 versions that Mistral says can run on four to eight Nvidia B200 or B300 GPUs. The company says ML4 was trained for two months on 4,000 Grace Blackwell GPUs in European data centers, supports more than 160 languages and will become the default model across its products after testing. Mistral reports strong results in grounding, cybersecurity, finance and industrial tasks, while acknowledging that some Chinese and closed models remain ahead on parts of agentic coding. Training and reinforcement learning were still underway, so the benchmark claims remain provisional and company-reported.
Hana’s perspective
I find the deployability claim more consequential than the benchmark race. Four to eight B200 or B300 GPUs is still expensive, but it puts a trillion-parameter model within reach of some studios, broadcasters and service providers that need scripts, visual references or production metadata to stay under their control. Open weights also permit specialized evaluation instead of accepting a vendor leaderboard. The caution is timing: ML4 is still training, its weights are not yet available, and the performance numbers come from Mistral. The real test will be whether independent users can reproduce the gains at the advertised hardware footprint and under workable licensing terms.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE