The model is described as having 1 trillion parameters, with 49 billion active, and as trained on Mistral’s cluster of 3,800 NVIDIA Grace Blackwell GPUs. The API offers two reasoning levels, “none” and “high.” Willison reports an Artificial Analysis score of 38 and compares it with a score of 9 for Mistral Large 3. Mistral says it plans to release open weights at the end of the month; that release remains prospective. The reported score is one benchmark measure, not a broad assessment of model quality or capability.
Raw SignalOct 6, 20261 min read1 source
Introducing Mistral Large 4: Le chonk
In one minute
Mistral has released a preview of Large 4 through its API, according to Simon Willison’s public account.