Magistral Medium 1.2 is a multimodal reasoning language model developed by Mistral AI, released in September 2025. It is part of the Magistral family, Mistral's dedicated line of reasoning-capable models designed to compete with frontier-class systems in complex problem-solving. While its smaller counterpart, Magistral Small, is released under open-weights licenses, the Medium variant is a proprietary enterprise model accessed via API or professional platforms.
Built for high-precision tasks, Magistral Medium 1.2 features native multimodality, allowing it to process and analyze both text and image inputs. The model is specifically optimized for multi-step logic, making it suitable for specialized domains such as legal research, financial forecasting, and advanced software development. It achieves significant performance gains over its predecessors, with reports indicating a 15% improvement in mathematics and coding benchmarks like AIME and LiveCodeBench.
To enhance transparency and interpretability, the model utilizes internal reasoning traces, often denoted by specific tokens such as [THINK] and [/THINK]. It supports a 128,000-token context window, providing the capacity to handle large documents and multi-turn conversations while maintaining deep reasoning performance. The model is also designed with improved tool-use capabilities, enabling more effective integration with external web search, code execution, and image generation tools.