On Tuesday, Mistral unveiled its inaugural artificial intelligence (AI) models designed for reasoning, named Magistral. This large language model (LLM) is available in two versions: Small and Medium. The Magistral Small model is open-source, while the Medium variant is tailored for enterprise use and operates on a closed model basis. Both versions possess the capability for multi-step reasoning and demonstrate a transparent chain-of-thought (CoT) process. The French AI startup noted that Magistral is crafted for applications in research, strategic planning, and data-driven decision making. Users can also experiment with a preview of the reasoning model on the company’s Le Chat platform.
Mistral’s Magistral Small Is Available With Apache 2.0 Licence
In an official announcement shared on its website, the Paris-based AI company detailed the launch of the two versions of Magistral. The open-weight Magistral Small can be downloaded and implemented via Mistral’s Hugging Face listing under the Apache 2.0 licence, allowing for both academic and commercial applications.
Conversely, the Magistral Medium model is proprietary and can be accessed through Amazon SageMaker. Plans are in place for it to be available on IBM WatsonX, Azure AI, and Google Cloud Marketplace in the near future. A preview version is also accessible on Le Chat and through the API on La Plateforme.
Regarding technical specifications, the Magistral Small features a model with 24 billion parameters, although the parameter count for the enterprise edition has not been disclosed by Mistral.
In terms of performance, Magistral Medium reportedly achieved a score of 73.6 percent on the AIME2024 benchmark, which is comparable to the DeepSeek-R1 model. In comparison, the Small version scored 70.7 percent, according to Mistral.
Both models exhibit native reasoning abilities, with the CoT functioning in several global languages, including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese.
Similar to many reasoning models, Magistral is equipped to handle structured calculations, programmatic logic, decision trees, and rule-based systems.
Mistral emphasized that users in critical sectors such as finance, healthcare, government, and law will benefit from traceable reasoning capabilities within the models. This feature will enable users to review the logical processes leading to specific responses, providing an important auditing tool for more sensitive inquiries.
