Mistral AI introduces a series of new language models focused on reasoning. This LLM series includes an open source and a commercial variant.
Mistral AI launches Magistral, its first reasoning model for language processing. The model is aimed at domain-specific and transparent reasoning in multiple languages. Magistral is available in two variants. An open-source variant (Magistral Small) with 24 billion parameters and a slightly more powerful commercial version (Magistral Medium).
Two variants
Magistral appears in two versions: Magistral Small (open-source) and Magistral Medium (commercial). Both are designed to support multi-step reasoning, suitable for applications requiring structured logic and clear thinking processes.
Magistral Small has 24 billion parameters and is available under an open license. Magistral Medium is more powerful and available through Amazon SageMaker, among others, and soon on IBM WatsonX, Azure AI, and Google Cloud Marketplace.
According to internal benchmarks, Magistral Medium achieved a score of 73.6 percent on the AIME2024 test and 90 percent in a voice test. Magistral Small scored 70.7 percent and 83.3 percent respectively.
The models support reasoning in multiple languages, including English, French, German, Spanish, Arabic, and Chinese.
Various sectors
Mistral AI positions Magistral for use in domains such as legal research, financial analysis, software development, data processing, and strategic planning. The model is said to be well-suited for regulated sectors such as healthcare, government, and finance, thanks to its ability to explain reasoning step by step and make it traceable.
read also
Mistral Small 3.1 Launched: Powerful Capabilities in a Compact Model
Magistral also works with Le Chat, an interactive interface from Mistral AI. The new “Flash Answers” feature in this tool is said to respond up to ten times faster than comparable models, enabling real-time feedback.
With the open-source release of Magistral Small, Mistral AI aims to invite the community to further develop the model. Previous models from the company have already been deployed for projects such as ether0 and DeepHermes 3.
Magistral is thus Mistral AI’s first model specifically designed to make reasoning language processing more transparent, reliable, and applicable in multilingual, business contexts.