Mistral AI has announced Magistral, its first reasoning language model designed for domain-specific, transparent, and multilingual thinking. The release includes two variants: Magistral Small, a 24B parameter open-source version under the Apache 2.0 license, and Magistral Medium, an enterprise-grade model available via API and cloud marketplaces.
- Magistral Medium scored 73.6% on AIME2024, reaching 90% with majority voting @64; Magistral Small achieved 70.7% and 83.3% respectively.
- The model supports native reasoning across global languages including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese.
- Le Chat users can access Flash Answers for up to 10x faster token throughput compared to most competitors.
- Magistral Small is available on Hugging Face, while Magistral Medium is accessible via La Plateforme, Amazon SageMaker, and upcoming integrations with IBM WatsonX, Azure AI, and Google Cloud Marketplace.
The release aims to provide traceable thought processes for high-stakes environments in regulated industries and complex engineering tasks, allowing users to verify conclusions through multi-step logic.