Back to glossary

Term

Mistral Large 3

Mistral Large 3 (December 2025) is Mistral's open-weight MoE flagship — 675B parameters, 256K-token context, natively multimodal, Apache 2.0.

Mistral Large 3 — explained in more detail

Mistral Large 3 was released on December 2, 2025 and is the current flagship from Mistral AI. The architecture is a Mixture of Experts with 675 billion total and 41 billion active parameters per token. A natively integrated vision encoder (~2.5B parameters) makes the model multimodal from the start — not a vision module bolted on later. This makes Large 3 the largest openly available MoE model from a major lab.

Specifications

  • Context window: 256K tokens — at the upper end among open models, relevant for RAG, code repositories, and long documents.
  • License: Apache 2.0 — commercial use permitted.
  • Multilingual: English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic.
  • Benchmarks: 73.1% on MMLU-Pro, 93.6% on MATH-500.

Context

Large 3 targets users who want open weights combined with top-tier quality — competitors in the open segment: DeepSeek V4, Llama 4 Maverick, Qwen 3.5. In March 2026, Mistral added the Forge platform for enterprise pre- and post-training on private datasets.

Entdecke mehr

Themenuebersicht