Magistral Small is a compact 24-billion-parameter reasoning model built on the Mistral Small 3.1 foundation, engineered to deliver strong chain-of-thought problem-solving within a lightweight footprint. Its architecture introduces special think tokens that encapsulate reasoning traces, making the model's thought process transparent and easy to parse. The model supports dozens of languages spanning English, French, German, Chinese, Arabic, and many more, while version 1.2 introduced a vision encoder for multimodal reasoning tasks. Designed for efficient local deployment, it fits within a single RTX 4090 or a 32GB RAM MacBook once quantized, bringing reasoning-class capabilities beyond the reach of larger, server-bound models.

The model climbs a deliberate training ladder: it begins with supervised fine-tuning on reasoning traces generated by its larger sibling, Magistral Medium, before receiving reinforcement learning to further sharpen its problem-solving approach. This ground-up pipeline, built entirely on Mistral's own models and infrastructure, demonstrates that reasoning ability can be cultivated through RL on text alone without sacrificing multimodal understanding or instruction-following. Benchmark results show solid performance across competition mathematics and reasoning benchmarks, with pass@1 scores reaching into the high 60s to low 70s depending on the task. Published under an Apache 2.0 license, it invites developers and researchers to inspect, modify, and deploy the model freely, making it a practical entry point for teams seeking an open-weight reasoning model that can be fine-tuned or run privately.