Mistral releases Magistral, its first reasoning models, with open-weight Magistral Small

Mistral AI/Mistral models

Mistral released Magistral, its first reasoning models: Magistral Small, a 24B open-weight model under Apache 2.0, and enterprise Magistral Medium. Medium scored 73.6% on the AIME 2024 maths exam (90% with majority voting) and reasons step by step in many languages; both support transparent, verifiable reasoning.

• Magistral Small: 24B parameters, open-weight under Apache 2.0; 70.7% AIME 2024 single-shot, 83.3% with voting.
• Magistral Medium: enterprise version; 73.6% AIME 2024 single-shot, 90% with majority voting (@64 attempts).
• 10x faster token throughput than competitors in Flash Answers mode.
• Both multilingual (8+ languages) with transparent, verifiable step-by-step reasoning.