Magistral Small 2506 is a compact reasoning-oriented language model from Mistral AI, built on the Mistral Small 3.1 2503 base and refined to think before it answers. With 24 billion parameters, it is designed to draft extended internal reasoning traces and then return a concise, useful reply, making it well suited for problems that benefit from step-by-step deliberation such as math, code, and multi-step analysis. The model supports a broad multilingual range spanning English, French, German, Spanish, Italian, Portuguese, Japanese, Korean, Chinese, Arabic, Hindi, and many more, and it carries an Apache 2.0 license that allows both commercial and non-commercial use. Its design intent is to bring stronger reasoning behavior into a smaller footprint that can run locally, fitting on a single RTX 4090 or a 32GB RAM MacBook once quantized.

The model's reasoning ability is layered on top of the base through supervised fine-tuning on traces from the larger Magistral Medium model, followed by additional reinforcement learning to sharpen its chain-of-thought behavior. This two-stage lineage lets a smaller 24B model inherit much of the reasoning style of its bigger sibling, and benchmark results reflect that transfer, with strong pass@1 scores on AIME24 and AIME25 alongside competitive results on GPQA Diamond and Livecodebench v5, sitting just below Magistral Medium. In practice, the model is a good fit for assistants, agentic workflows, coding help, and multilingual chat where structured thinking and tool use matter, and it pairs well with inference stacks like vLLM and quantized runtimes such as llama.cpp, LM Studio, Ollama, and Unsloth. Looking ahead, its combination of open licensing, local deployability, and reasoning-first tuning positions it as a practical middle ground between lightweight chat models and heavier frontier systems.