Magistral Small and Medium
Magistral is Mistral's first reasoning model. Medium scores 73.6% on AIME 2024 (90% with majority voting), and Small is open under Apache 2.0.
Built with Mistral's own scalable RL pipeline, no distilled traces from rival models. Paper shows RL on text alone keeps or improves multimodal and instruction-following ability. Reasoning in many languages (English, French, Spanish, German, Arabic, more).
- Date
- Tuesday, 10 June 2025
- Lab
- Mistral AI
- Kind
- model
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2024 (Medium) | 73.6% 90% with majority voting | company |
| AIME 2024 (Small) | 70.7% 83.3% with majority voting | company |
Magistral Small open weights; Medium via API and Le Chat. Paper arXiv 2506.10910 (v1 2025-06-12, 99 authors). Mistral was about five to nine months behind OpenAI o1 and DeepSeek-R1 (inference from dates).
Sources
This record was checked against its sources on 6 October 2026. How we check