TechPrufer
Sign inGet started

Blogs›Mistral AI›Magistral Reasoning Model

Magistral Reasoning Model

Magistral Reasoning Model

2
posts
2024–2025

Mistral AI introduces Magistral, a new family of reasoning models designed for domain-specific, transparent, and multilingual reasoning. The models are released in two variants: Magistral Small (24B parameters, open-source) and Magistral Medium (enterprise version). Magistral excels in multi-step logic, providing traceable thought processes and high-fidelity reasoning across multiple languages. It is optimized for speed, with Magistral Medium achieving up to 10x faster token throughput in Le Chat. This post introduces MathΣtral, a specialized 7B parameter model derived from Mistral 7B, focusing on advanced mathematical problems requiring complex, multi-step logical reasoning. MathΣtral achieves state-of-the-art reasoning capacities in its size category across industry-standard benchmarks like MATH and MMLU, demonstrating significant performance improvements over its base model in STEM subjects. It can achieve even better results with increased inference-time computation, such as majority voting or using a strong reward model.

2025

Magistral | Mistral AI

6/10/2025

This post announces Magistral, Mistral AI's first reasoning model. It details two variants: Magistral Small (open-source, 24B parameters) and Magistral Medium (enterprise). Performance metrics on AIME2024 are provided. Key features highlighted include native multilingual chain-of-thought reasoning, suitability for enterprise use cases like structured calculations and decision trees, and significantly faster reasoning speeds (10x throughput) via 'Flash Answers' in Le Chat. The post also references a supporting research paper detailing training infrastructure, reinforcement learning algorithms, and novel observations for training reasoning models. Availability details for both open-weight and enterprise versions are provided, along with information on integrations with cloud platforms.

2024

MathΣtral | Mistral AI

7/16/2024

Introduces MathΣtral, a specialized 7B parameter model derived from Mistral 7B, focusing on advanced mathematical problems requiring complex, multi-step logical reasoning. MathΣtral achieves state-of-the-art reasoning capacities in its size category across industry-standard benchmarks like MATH (56.6%) and MMLU (63.47%), demonstrating significant performance improvements over its base model in STEM subjects. It can achieve even better results with increased inference-time computation, such as 68.37% on MATH with majority voting and 74.59% with a strong reward model among 64 candidates. The model is released under an Apache 2.0 license and is available via HuggingFace, mistral-inference, and mistral-finetune.