Mistral drops unified 119B MoE reasoning model

Mistral's latest release proves that MoE scaling can deliver elite reasoning without the massive compute tax of dense models.

• –The model consolidates three previously distinct branches into one, simplifying workflows for developers who need both speed and deep logic.
• –Speculative decoding via an "eagle head" and NVFP4 support significantly reduce end-to-end completion time by 40% compared to previous versions.
• –Matches or beats GPT-OSS 120B and Qwen on major benchmarks while maintaining much shorter, more concise outputs.
• –Native vision support and specialized agentic capabilities like tool-use and structured JSON output make it a premier choice for complex automation.
• –Apache 2.0 license is a massive win for the open-source community, providing a top-tier alternative to proprietary reasoning models.