Introducing Mistral 3

Summary

Mistral AI's announcement of the Mistral 3 family, led by Mistral Large 3, a sparse mixture-of-experts model, with every model released under Apache 2.0. Details here come from Mistral's announcement and its Hugging Face card as quoted in search results.

Key points

• Mistral Large 3: about 675B total and 41B active parameters, quoted as 673B/39B for the language model plus a 2.5B vision encoder.
• 256K-token context window.
• Multilingual, including the main European languages, Chinese, Japanese, Korean, and Arabic.
• Native function calling and JSON output for agent use.
• All Mistral 3 models are licensed under Apache 2.0.

Why it matters

The most capable fully permissive open-weight model from a European lab, often chosen when data must stay in EU deployments.