Magistral

Magistral is Mistral AI's first reasoning model family, released in June 2025 in two variants: Magistral Small (24B parameters, open weights under Apache 2.0) and Magistral Medium (enterprise tier). It reasons step by step with visible, traceable chains of thought, and its standout feature is multilingual reasoning - it can think natively in French, Spanish, German, Italian, Arabic, Russian, Simplified Chinese, and English. An updated Magistral Small 1.2 later added image analysis and runs on consumer hardware. Honest limitation: on release-day benchmarks it trailed frontier peers like Gemini 2.5 Pro and Claude Opus 4 on GPQA Diamond and LiveCodeBench, and Mistral is explicit that it iterates the model continuously.

AIME2024 (from Mistral's official blog): Magistral Medium 73.6% (90% with majority vote @64); Magistral Small 70.7% (83.3% with majority vote @64). TechCrunch reported that on Mistral's own charts, Medium underperformed Gemini 2.5 Pro and Claude Opus 4 on GPQA Diamond and LiveCodeBench - no exact scores published for those.