On June 10, 2025, Mistral AI released Magistral, the company’s first reasoning model, in two versions. Magistral Small is a 24-billion-parameter open-weight model released under the Apache 2.0 license on Hugging Face, which Mistral said the community can “examine, modify and build upon.” Magistral Medium is a more powerful proprietary version available through Mistral’s Le Chat assistant and La Plateforme API, with planned availability on Amazon SageMaker, IBM WatsonX, Azure AI, and Google Cloud.
Mistral emphasized transparent, multilingual reasoning aimed at regulated industries, saying “Magistral’s chain-of-thought works across global languages and alphabets” and provides “traceable reasoning that meets compliance requirements.” On the AIME2024 math benchmark, Mistral reported 70.7 percent for Magistral Small and 73.6 percent for Magistral Medium, rising to 83.3 and 90 percent respectively with majority voting over 64 samples, and claimed Magistral Medium delivers “up to 10x faster token throughput than most competitors.”
Magistral mattered because it brought a European lab into the reasoning-model race with an explicit focus on auditable, multilingual chains of thought. Pairing an open small model with a proprietary enterprise model, it positioned reasoning as a compliance and transparency feature rather than just a benchmark gain.