In 2025, the French AI startup Mistral AI rolled out its newest open-source reasoning model, named Magistral Small 1.2. This cutting-edge model comes packed with a staggering 24 billion parameters and is made available under the widely recognized Apache 2.0 license.

One of its standout features is its ability to handle context processing of up to 128k, which is a substantial leap in terms of processing capacity. Moreover, it boasts multilingual processing capabilities, allowing it to seamlessly work with various languages. Additionally, it can process visual inputs, making it a versatile tool in the AI landscape.

Magistral Small 1.2 takes a step further by introducing a unique [THINK] token. This innovative addition significantly boosts the model's expressiveness and flexibility, enabling it to tackle complex reasoning tasks with greater ease.

Furthermore, the model incorporates a brand-new visual encoder. This encoder equips Magistral Small 1.2 with a competitive advantage when it comes to handling tasks that involve the integration of images and text, such as image captioning or visual question answering.