Mistral AI Masterclass: The Reasoning Model That Accelerates Complex Analysis | | Llm training dataset github | Llm dataset example | Turtles AI
MistralAI’s Magistral is an innovative reasoning model, open-source in the Small version, and enterprise in the Medium version. It excels in multilingual performance, transparency, speed, multi-step capability, with AIME2024 scores between 70.7% and 73.6%, 90% with voting.
Key points:
- Dual version: Small (24B parameters, open-source Apache2.0), Medium (enterprise).
- AIME2024 scores: Medium 73.6% (90% with voting), Small 70.7% (83.3%).
- Transparent “chain-of-thought” reasoning in multiple languages (including Italian, Arabic, Simplified Chinese).
- Up to 10× speed thanks to “Flash Answers” in Think Mode on LeChat.
The new Magistral, unveiled on June 10, 2025, is the first reasoning model developed by Mistral AI: available in an open-source version (Magistral Small, 24 billion parameters, Apache 2.0 license) and in an enterprise version (Magistral Medium), it stands out for its ability to tackle multi-step problems with transparent reasoning and verifiable supervision, reaching performances between 70.7% and 73.6% on AIME2024, with up to 90% if supported by a vote on 64 answers. Thanks to the Think mode and the Flash Answers functionality within the LeChat interface, the Medium model reaches a token processing speed up to ten times faster than most competitors, while maintaining high accuracy. Magistral also stands out for its linguistic versatility: it supports coherent reasoning in several languages, including English, French, Spanish, German, Italian, Arabic, Russian and Chinese, overcoming the limitations of typically monolingual models. Technically, the training of Magistral Medium is based on an innovative Reinforcement Learning from Verifiable Rewards (RLVR) pipeline, implementing Group Relative Policy Optimization to qualitatively improve the multi-step logical reasoning, while the Small derives from a fine-tuning on reasoning traces generated by the Medium. The architecture and parameters are detailed in the official paper, which includes analysis on the training infrastructure, enablement of the reasoning chain-of-thought and preservation of the original multimodal capabilities. Magistral is suitable for specialized business scenarios: from strategic planning and risk models to compliance in regulated sectors, up to structured software engineering, thanks to the traceability of deductions. The open-weight version is already available on Hugging Face, ready for download and integration, while the Medium version is accessible via LeChat, API on La Plateforme and cloud platforms (Amazon SageMaker, IBM WatsonX, Azure and Google Cloud).
Magistral marks a new step towards advanced reasoning models, with transparency, multilingualism and concrete use in the most demanding contexts.
