Magistral 1.2
2025-09-17 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI releases Magistral Medium 1.2 and Magistral Small 1.2, which add a vision encoder (multimodal reasoning) and significantly improve reasoning scores, with Small remaining under the Apache 2.0 license.
What changes
- Multimodality: a vision encoder makes it possible to reason over images in addition to text.
- Clear performance gain over 1.1: AIME 2025 from 60.99% to 83.48% for Medium.
- Special tokens [THINK] and [/THINK] enclosing the reasoning, to make the trace easier to analyze.
- Follows Magistral 1.1 (July 24, 2025), a minor update with performance comparable to 1.0.
Mistral AI releases Magistral Medium 1.2 and Magistral Small 1.2, which add a vision encoder (multimodal reasoning) and significantly improve reasoning scores, with Small remaining under the Apache 2.0 license. ### What changes - Multimodality: a vision encoder makes it possible to reason over images in addition to text. - Clear performance gain over 1.1: AIME 2025 from 60.99% to 83.48% for Medium. - Special tokens [THINK] and [/THINK] enclosing the reasoning, to make the trace easier to analyze. - Follows Magistral 1.1 (July 24, 2025), a minor update with performance comparable to 1.0. ### Specifications - Magistral Small 1.2: 24 billion parameters, Apache 2.0 license, weights on Hugging Face; Magistral Medium 1.2: proprietary. - Context: 128k (the HF card indicates possible degradation beyond 40k). - API identifiers: `magistral-medium-2509`, `magistral-small-2509` (Medium 1.2 deprecated on 05/22/2026 according to the documentation). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Magistral Medium | 2 | 5 | | Magistral Small | 0.5 | 1.5 | Prices from the pricing page archived on July 30, 2025 (versions 1.0 and 1.1); not verified for version 1.2. ### Benchmarks Scores published by the vendor (Hugging Face card): | Benchmark | Medium 1.2 | Small 1.2 | Medium 1.1 | |---|---|---|---| | AIME 2024 (pass@1) | 91.82% | 86.14% | 72.03% | | AIME 2025 (pass@1) | 83.48% | 77.34% | 60.99% | | GPQA Diamond | 76.26% | 70.07% | 71.46% | | LiveCodeBench (v5) | 75.00% | 70.88% | 59.35% | ### Key takeaways A major evolution of the reasoning line (vision and accuracy gains), available as an open version for internal deployment.
Links
Tags : LLM, Mistral, Magistral, Raisonnement, Multimodal, Open Source