Magistral Small 1.2 is a 24-billion parameter multimodal reasoning model developed by Mistral AI. Released in September 2025, it represents an update to the Magistral family, specifically succeeding version 1.1 and building upon the foundation of the Mistral Small 3.2 architecture. The model is optimized for efficiency, designed to provide advanced intelligence while remaining runnable on consumer-grade hardware.
A primary feature of Magistral Small 1.2 is its integrated vision encoder, which enables native multimodal reasoning across both text and image inputs. To support complex problem-solving, the model utilizes a specialized reasoning format with [THINK] and [/THINK] tokens. these tokens encapsulate internal chain-of-thought processes, allowing the model to perform deep analysis before generating a final response without exposing the raw trace to the end user unless desired.
The model features a 128,000-token context window and provides significant performance improvements on reasoning-heavy benchmarks such as AIME and GPQA. It is highly multilingual, supporting a wide array of languages including English, French, German, Spanish, Italian, Chinese, Japanese, and Arabic. Magistral Small 1.2 is released under the Apache 2.0 license, permitting both commercial and non-commercial utilization.