Mistral releases a pair of AI reasoning models

Summary

French AI lab Mistral is getting into the reasoning AI model game. On Tuesday morning, Mistral announced Magistral, its first family of reasoning models. Like other reasoning models — e.g. OpenAI’s o3 and Google’s Gemini 2.5 Pro — Magistral works through problems step-by-step for improved consistency and reliability across topics such as math and physics. Magistral comes in two flavors: Magistral Small and Magistral Medium. Magistral Small is 24 billion parameters in size, and is available for download from the AI dev platform Hugging Face under a permissive Apache 2.0 license. (Parameters are the internal components of a model that guide its behavior.) Magistral Medium, a more capable model, is in preview on Mistral’s Le Chat chatbot platform and the company’s API, as well as third-party partner clouds. “[Magistral is] suited for a wide range of enterprise use cases, from structured calculations and programmatic logic to decision trees and rule-based systems,” writes Mistral in a blog post. “[The models are] fine-tuned for multi-step logic, improving interpretability and providing a traceable thought process in the user’s language.” Founded in 2023, Mistral is a frontier model lab building a range of AI-powered services, including the aforementioned Le Chat and mobile apps. It’s backed by venture investors like General Catalyst, and has raised over €1.1 billion (roughly $1.24 billion) to date. Despite its formidable resources, Mistral has lagged behind other leading AI labs in certain areas, like developing reasoning models. Magistral doesn’t appear to be an especially competitive release, either, judging by Mistral’s own benchmarks. On GPQA Diamond and AIME, tests that evaluate a model’s physics, math, and science skills, Magistral Medium underperforms Gemini 2.5 Pro and Anthropic’s Claude Opus 4. Magistral Medium also fails to surpass Gemini 2.5 Pro on a popular programming benchmark, LiveCodeBench. Perhaps that’s why Mistral touts Magistral’s other strengths in ...