Magistral is Mistral AI's inaugural reasoning model, designed for domain-specific, transparent, and multilingual problem-solving. It is released in two variants: Magistral Small, a 24B parameter open-source version, and Magistral Medium, a more powerful enterprise version. Performance benchmarks on AIME2024 demonstrate Magistral Medium achieving 73.6% (90% with majority voting @64) and Magistral Small scoring 70.7% (83.3% respectively).

The core methodology of Magistral emphasizes native, multilingual chain-of-thought (CoT) reasoning, enabling it to process and articulate logical steps across diverse languages and alphabets. This capability is inherent to the model's design, rather than relying on external translation or post-processing. The model is specifically fine-tuned for multi-step logic, a process that enhances interpretability by generating a traceable thought process. This fine-tuning likely involves supervised learning on datasets annotated with detailed reasoning steps, potentially followed by reinforcement learning to refine the quality and coherence of these steps, as hinted by the mention of a "reinforcement learning algorithm" in the supporting technical paper. The goal is to produce explicit intermediate steps, allowing users to follow and verify the model's conclusions, a critical feature for transparency and audibility in regulated industries.

Magistral's architecture and training infrastructure are optimized for reasoning tasks, suggesting specialized design choices for handling complex dependencies and propagating information across multi-step computations. While specific architectural details are not provided, the focus on "novel observations for training reasoning models" implies a departure from or enhancement of standard large language model pre-training and fine-tuning regimes. The integration of "Think mode" and "Flash Answers" in Le Chat indicates significant inference optimizations, achieving up to 10x faster token throughput than competitors. This speed improvement could stem from highly optimized inference kernels, selective attention mechanisms, or efficient decoding strategies designed to quickly produce high-quality reasoning outputs.

The model demonstrates strong multilingual dexterity, maintaining high-fidelity reasoning in languages such as English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. This suggests that the training data and fine-tuning processes are deeply multilingual, allowing the model to perform complex logical operations directly within various linguistic contexts without performance degradation. Magistral is engineered for versatility, excelling in applications requiring precise, multi-step deliberation, including structured calculations, programmatic logic, decision trees, risk assessment, legal research, financial forecasting, software development (project planning, backend/frontend, data engineering through sequenced actions and external tool/API integration), and creative content generation.

Magistral Small is available as an open-weight model under the Apache 2.0 license, promoting community examination and development. Magistral Medium is accessible via Mistral AI's Le Chat preview, La Plateforme API, Amazon SageMaker, and soon on IBM WatsonX, Azure AI, and Google Cloud Marketplace.