Mistral Large 3

Mistral Large 3 is the flagship release in Mistral AI's Mistral 3 lineup, positioned as the company's most capable open-weight model to date. It is released under the Apache 2.0 license alongside smaller dense Ministral variants, signaling Mistral's continued commitment to permissive open distribution. The model is officially documented with version tag v25.12 and identified by the path mistral-large-3-25-12, and it is marked as generally available for inference use.

At the architectural level, Mistral Large 3 is a sparse mixture-of-experts model with 41 billion active parameters out of 675 billion total, trained from scratch on a cluster of 3000 NVIDIA H200 GPUs. This is Mistral's first MoE design since the Mixtral series and represents a substantial step forward in pretraining at the company. After post-training it achieves parity with leading instruction-tuned open-weight models on general prompts while also demonstrating image understanding, making it a strong fit for general-purpose multimodal workloads that benefit from open weights.