Simon Willison
Mistral Previews Mistral Large 4: A 1T-Parameter MoE with Upcoming Open Weights
Original title:Introducing Mistral Large 4: Le chonk
Models72
Mistral has previewed Mistral Large 4, an MoE model featuring 1 trillion total parameters and 49 billion active parameters trained on 3,800 NVIDIA Grace Blackwell GPUs. Currently accessible through their API with two reasoning settings, Mistral plans to release the open-weight model later this month. Early benchmark indicators suggest Mistral is steadily closing its gap to the frontier.
Why it's worth reading
Mistral's return with a 1T-parameter MoE architecture trained on custom Blackwell clusters demonstrates how open-weight developers attempt to close the distance to frontier labs.
Tags
MistralMistral Large 4MoEOpen WeightsNVIDIA BlackwellLLM