Mistral AI Announces Mistral Large 4 'Le Chonk' with Open Weights Coming Soon
Mistral AI has announced Mistral Large 4, codenamed 'Le Chonk,' featuring a massive 1-trillion parameter architecture. The model relies on a Mixture-of-Experts system with 49 billion active parameters, with open weights expected to be released by the end of the month. This release reinforces Europe's footprint in frontier AI development and brings unprecedented parameter scale to the open-weights community. By making a 1-trillion parameter MoE model open-weight, Mistral AI enables developers and researchers to run frontier-class models with manageable active inference costs. While the total footprint reaches 1 trillion parameters, only 49 billion active parameters are triggered for any given token during inference. This design provides deep domain capacity across experts while keeping compute latency and hardware bandwidth constraints significantly lower than a dense model of equivalent total size.
## BACKGROUND
A Mixture-of-Experts (MoE) architecture divides an AI network into multiple specialized subnetworks called experts, relying on a router to send input tokens only to the relevant experts. This approach differentiates total parameters (the overall network size stored in memory) from active parameters (the specific parameters used per token pass). As a result, MoE models achieve state-of-the-art capabilities without requiring the compute footprint of a fully active dense network.