Magistral
By Mistral AI
Magistral is Mistral AI's family of reasoning-focused large language models, released in 2025, designed to produce explicit step-by-step chain-of-thought reasoning traces for improved performance on complex math, logic, and multi-step…
Definition
Magistral is Mistral AI's family of reasoning-focused large language models, released in 2025, designed to produce explicit step-by-step chain-of-thought reasoning traces for improved performance on complex math, logic, and multi-step problem-solving tasks.
Overview
Magistral is Mistral AI's entry into the 'reasoning model' category that became prominent following OpenAI's o1 and subsequent reasoning-focused releases from other labs (such as DeepSeek-R1 and Google's Gemini thinking models). Rather than producing a direct answer immediately, Magistral models are trained to generate an extended internal chain-of-thought — working through a problem step by step, checking intermediate results, and backtracking when necessary — before committing to a final answer, which tends to substantially improve accuracy on tasks requiring multi-step logical, mathematical, or symbolic reasoning compared to models optimized purely for fast single-pass responses. Mistral released two variants: Magistral Small, an open-weight model made available under the Apache 2.0 license for self-hosting and research, and Magistral Medium, a more capable proprietary model accessible via Mistral's API. This mirrors Mistral's general pattern of pairing a smaller open release with a larger proprietary flagship. Magistral was noted at release for supporting reasoning traces across multiple languages, not just English, and for maintaining relatively transparent, readable chain-of-thought outputs, which Mistral positioned as valuable for auditability in domains like legal or scientific analysis where understanding the model's reasoning path matters as much as the final answer. Magistral fits into the broader 2024-2025 trend of 'test-time compute' or 'inference-time scaling' models, where labs found that letting a model spend more computation generating intermediate reasoning steps before answering — rather than only scaling up pretraining — could meaningfully improve performance on hard reasoning benchmarks like competition mathematics and complex coding problems, at the cost of higher latency and token usage per query.
Key Concepts
- Generates explicit step-by-step chain-of-thought reasoning before final answers
- Two variants: open-weight Magistral Small (Apache 2.0) and proprietary Magistral Medium
- Designed for improved accuracy on math, logic, and multi-step reasoning tasks
- Supports multilingual reasoning traces, not just English
- Emphasizes relatively transparent, readable chain-of-thought for auditability
- Part of the broader 'test-time compute' / inference-time scaling model category