Mistral AI, a French company, has recently officially launched its latest open-source reasoning model - Magistral Small1.2. This model has 24B parameters and is released under the Apache 2.0 open-source license, fully demonstrating Mistral's ambition and strength in AI technology innovation.
The new version, Magistral Small1.2, not only supports up to 128k context processing, can handle multiple languages and visual inputs, but also introduces a new innovative feature - [THINK] special token. This special token is designed to wrap the reasoning process, thereby enhancing the model's expressiveness and flexibility. Compared to the previous version 1.1, Magistral Small1.2 has added a visual encoder, making it more advantageous in handling integrated tasks of images and text.
In addition, Magistral Small1.2 comes with multiple reasoning templates, compatible with popular frameworks such as vLLM, Transformers, and llama.cpp, allowing users to get started immediately without complicated configurations. At the same time, the model also provides GGUF quantization versions and Unsloth fine-tuning examples, offering more convenience for developers.
The enterprise version of Magistral Medium1.2 has also been upgraded synchronously. It continues to provide dialogue services through the Le Chat platform, and its API has been launched on the La Plateforme platform, further expanding application scenarios.
Mistral's series of new initiatives mark an important step forward in the open-source AI field, providing developers and enterprises with more powerful and flexible tools to help them succeed in the rapidly developing AI market. Whether researchers or enterprise users, they can leverage this new model to improve their work efficiency and technical level.