Magistral Small 2506

Magistral Small 2506 is a 24-billion-parameter language model developed by Mistral AI as part of its first reasoning-focused family. It is built on the Mistral Small 3 base, extending that lineage with added reasoning capabilities and multilingual support across dozens of languages. Within the Magistral lineup, this variant ships as the open-weights, Apache 2 licensed option, while a larger sibling called Magistral Medium is offered as an API-only model. The Hugging Face repository is mistralai/Magistral-Small-2506, and Mistral also publishes an official GGUF quantization for local runtimes.

According to the Magistral-Small-2506 model card, the model advertises a 128k context window but recommends capping practical usage at around 40k tokens, where performance remains reliable before degradation may set in. The model is aimed at reasoning-heavy workloads, fast-response conversational agents, and applications that combine strong language understanding with structured problem solving. Its blend of a relatively compact parameter count, open licensing, and dedicated reasoning training makes it a practical choice for teams that want Mistral's chain-of-thought style generation without committing to a larger proprietary deployment.