Mistral Large 4 is available to test through an API, but businesses cannot yet download its weights or run the model on their own systems.

French AI company Mistral opened the public API preview on October 6. It says it plans to publish the model weights by the end of October. Until then, teams can try Large 4 through Mistral’s service, but cannot deploy or adapt the model themselves.

Testing the model is not the same as running it

Mistral describes Large 4 as natively multimodal. The company reports that it was trained on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.

  • 1 trillion total parameters
  • 49 billion active parameters
  • 3,800 NVIDIA Grace Blackwell GPUs used in training, according to Mistral

Model weights are the files needed to run or adapt a model on a user’s own hardware. TechCrunch reported that the current preview uses a monitored endpoint while Mistral completes safety testing. API access therefore lets businesses evaluate the model, but leaves deployment in Mistral’s hands for now.

Early scores are not an independent verdict

Mistral says Large 4 is competitive in coding, finance, cybersecurity and other specialist tasks. The company reported a 63% score on Deep SWE 1.1, a benchmark for long coding tasks. That result is a company-reported benchmark, not proof that Large 4 leads across everyday uses.

Le Monde reported that the model’s performance had not yet been confirmed in regularly updated independent rankings. For developers and businesses, the practical test is whether the API handles their own work, and whether later results hold up in independent comparisons.

Weights are planned for October 27

Mistral says it plans to publish the weights and further benchmarks by the end of October. Le Monde reported October 27 as the planned public release date. Until then, teams can test Large 4 through the API, but cannot inspect or deploy its weights themselves.