Model Releases
Magistral Medium 1.2: Vision-Enhanced Reasoning Power Unveiled
Mistral AI releases Magistral Medium 1.2, adding vision capabilities to its 45B parameter reasoning engine with a closed API model.
September 1, 2025
Model ReleaseMagistral Medium 1.2
Introduction
Mistral AI has officially unveiled Magistral Medium 1.2 on September 1, 2025, marking a significant evolution in their reasoning model lineup. This release is not merely an incremental update but a strategic leap towards multimodal frontier reasoning, addressing the critical gap between language understanding and visual comprehension. For developers building complex AI agents, this model represents a consolidation of high-level logic with visual perception.
The decision to keep Magistral Medium 1.2 closed-source underscores Mistral's focus on premium enterprise-grade reliability rather than community iteration. With the release, Mistral continues to solidify its position as Europe's leading AI challenger, complemented by their recent €1.2 billion investment in Swedish infrastructure to ensure sovereign compute capacity. This model is designed for scenarios where reasoning requires visual context, such as diagram analysis or code visualization.
- Release Date: September 1, 2025
- Architecture: Closed API Only
- Focus: Multimodal Reasoning
Key Features & Architecture
Magistral Medium 1.2 operates with approximately 45 billion parameters, utilizing a sophisticated Mixture of Experts (MoE) architecture to optimize inference speed without sacrificing reasoning depth. The model features a native context window of 128,000 tokens, allowing it to process long-form documents alongside visual inputs seamlessly. Unlike previous iterations, this version integrates a high-resolution vision encoder directly into the transformer backbone, enabling true multimodal reasoning rather than simple image captioning.
The integration of vision capabilities allows the model to interpret charts, graphs, and code snippets with significantly higher accuracy. This architecture supports dynamic tokenization for visual data, ensuring that complex mathematical derivations shown in images are understood correctly. The closed nature of the API ensures that the proprietary vision-language alignment remains secure for enterprise clients requiring strict data governance.
- Parameters: ~45 Billion
- Context Window: 128k Tokens
- Vision: Native Multimodal Encoder
- Architecture: MoE (Mixture of Experts)
Performance & Benchmarks
In terms of performance, Magistral Medium 1.2 demonstrates substantial improvements over the previous Medium 1.0 variant. On the MMLU benchmark, it achieves a score of 84.5, surpassing the previous 82.1 mark. For developers evaluating coding capabilities, the HumanEval score reaches 88.2, indicating strong proficiency in generating and debugging Python functions. The SWE-bench leaderboard shows a 15% improvement in solving real-world software issues compared to the baseline.