Tune into our latest podcast episode where we delve into Magistral, Mistral AI's groundbreaking first reasoning model. This innovative system pioneers a scalable reinforcement learning (RL) pipeline, built entirely from the ground up without relying on existing distilled traces. A key novelty is its demonstration of pure RL training for large language models (LLMs), showing that RL on text data alone significantly boosts capabilities like multimodal understanding, instruction following, and function calling. For instance, Magistral Medium, trained solely with RL, achieved a nearly 50% increase in AIME-24 accuracy over its base model.

While powerful, the model experiences a slight degradation in multilingual reasoning compared to English, performing 4.3-9.9% lower on AIME 2024 benchmarks. Additionally, experiments with proportional rewards for code tasks or entropy bonuses for exploration were unsuccessful or unstable, suggesting nuances in RL application. Magistral is primarily applied to complex mathematical and coding problems, with proven efficacy in multilingual and multimodal contexts. Its strong foundation supports future advancements in tool-use and intelligent agents.

Find the full paper here: https://arxiv.org/pdf/2506.10910