Mistral releases Magistral, its first reasoning model
Mistral launched Magistral Medium (a preview, via Le Chat and the API) and the open Magistral Small (24B, Apache 2.0) on 2025-06-10, 271 days after o1-preview and 141 after R1.
- Date
- 10 June 2025
- Who
- Mistral AI
- Confidence
- High
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): Mistral AI · Confidence: High Mistral launched Magistral Medium (a preview, via Le Chat and the API) and the open Magistral Small (24B, Apache 2.0) on 2025-06-10, 271 days after o1-preview and 141 after R1. Its own announcement calls Magistral "the first reasoning model by Mistral AI" and makes no Europe-wide claim. Mistral reported AIME 2024 of 73.6% for Medium (90% with majority voting at 64 samples) and 70.7% (83.3%) for Small, and claimed strong multilingual reasoning (blog). The paper (arXiv 2506.10910, 2025-06-12) says Mistral built its own RL stack and did not distill from existing reasoning models, found that RL on text alone preserved multimodal ability, instruction-following and function calling, and gave Small a cold start with traces from the larger model (arXiv). It matters here because it is among the slower established-lab replications (Meta's first flagship reasoning model came 573 days after o1-preview; its earlier small MobileLLM-R1 models, B08-37a, 365). I infer that diffusion depended on engineering capacity for RL infrastructure and did not depend on secret knowledge. See also B19.