Mistral Launches Magistral AI Models with Open Source and Enterprise Focus

Mistral, a leading European AI company, has launched Magistral, a new family of large language models (LLMs) that focus on advanced reasoning capabilities. This release marks a significant step as these models are designed to reflect on their thinking process, catch errors, and solve complex tasks beyond basic text generation.

Magistral comes in two variants: Magistral Medium, a proprietary model tailored for enterprise clients, and Magistral Small, a 24-billion parameter open-source model released under the permissive Apache 2.0 license. This open-source approach empowers developers and businesses to freely use, modify, and distribute the model, fostering innovation without licensing constraints.

This dual release strategy addresses concerns from the AI community about Mistral’s previous shift towards proprietary models, reaffirming its commitment to open-source roots while offering powerful enterprise-grade solutions.

Performance benchmarks demonstrate Magistral Medium’s competitive edge. On the AIME-24 mathematics test, it achieves 73.6% accuracy, rivaling its predecessor and outperforming competitors like Deepseek. Using majority voting, accuracy jumps to 90%. It also excels in graduate-level question answering and coding challenges, proving its versatility and strength in reasoning.

Magistral Medium is accessible via Mistral’s Le Chat interface and La Plateforme API, with availability on major cloud platforms like Amazon SageMaker and upcoming support on Azure AI, IBM WatsonX, and Google Cloud Marketplace. This makes it a practical choice for enterprises requiring reliable, scalable AI reasoning.

Pricing positions Magistral Medium as a premium offering at $2 per million input tokens and $5 per million output tokens. While more expensive than Mistral’s older models, it is competitively priced against rivals like OpenAI and Gemini 2.5 Pro, offering a compelling balance of cost and performance.

Magistral emphasizes three core advantages: transparency, multilingualism, and speed. It produces a traceable chain-of-thought, enabling users to follow the model’s reasoning — a critical feature for sectors like law, finance, and healthcare. It supports multiple languages including French, Spanish, German, Arabic, Russian, and Chinese, broadening its global applicability.

Performance enhancements include a “Think mode” and “Flash Answers” feature that boost token throughput by up to 10 times compared to competitors, enabling real-time reasoning at scale.

The applications for Magistral are wide-ranging, from financial modeling and legal analysis to software architecture and data engineering. It can even generate complex physics simulations and serve as a creative companion for writing and storytelling, showcasing both precision and creative flexibility.

By combining open-source accessibility with enterprise-grade performance, Mistral’s Magistral models represent a strategic move to lead the future of AI reasoning — powerful, transparent, and open to all.