Mistral Large 3

Mistral Large 3 is Mistral AI's flagship released December 2025: a sparse mixture-of-experts with 675B total and 41B active parameters, multimodal (text and image input), a 256K context window, and multilingual coverage across 40+ languages. It placed near the top of the open-weight field on LMArena at launch, scored 1418 Elo in LMSYS testing, and hit 92.3% on HumanEval. Trained from scratch on 3,000 NVIDIA H200 GPUs, it marked a break from Mistral's earlier closed releases by shipping under Apache 2.0 on day one.

The honest picture: 675B total parameters means self-hosting is a multi-GPU-node affair despite the sparse active count; practical users rely on the Le Chat/API or quantized builds. Like most generalist frontier models it trades away extreme multi-step reasoning depth (mid-40s on GPQA Diamond) in favor of broad capability and throughput. The Mistral 3 family announcement was accompanied by heavy funding narratives, so some benchmark framing should be read with vendor-report caution.