Mistral AI Previews Mistral Large 4, a 1-Trillion Parameter MoE Reasoning Model
Mistral AI has announced an API preview for Mistral Large 4, a 1-trillion total parameter Mixture of Experts (MoE) reasoning model trained on a cluster of 3,800 NVIDIA Grace Blackwell GPUs. The company plans to release the model with open weights by the end of the month. This launch represents a major leap forward for Mistral AI, closing the gap with top frontier models while reinforcing their commitment to open-weight foundation models. It gives developers and researchers upcoming access to powerful, large-scale reasoning capabilities outside walled AI ecosystems. While the model boasts 1 trillion total parameters, it only activates 49 billion active parameters per token during inference, and its API currently supports two reasoning effort levels ('none' and 'high'). On Artificial Analysis benchmarks, Mistral Large 4 scored 38, showing a dramatic improvement over Mistral Large 3's score of 9.
## BACKGROUND
Mixture of Experts (MoE) is a neural network architecture that splits parameters into multiple specialized expert sub-networks, activating only a small subset per token to keep inference computation low while maintaining massive total capacity. Reasoning models build upon standard language models by incorporating step-by-step thinking processes (such as chain-of-thought) to solve complex analytical or coding tasks before generating final outputs.