Mistral has unveiled a hybrid, multimodal model alongside a new platform that lets enterprises build custom models trained on their own data.
Mistral AI’s newest model in the fully open-source ‘Small’ series will attempt to consolidate capabilities of its flagship models, according to the company.
The ‘Mistral Small 4’ is a hybrid model optimised for a variety of tasks, the French company said, powered with reasoning strengths from the ‘Magistral’ model, multimodal capabilities from ‘Pixtral’ and coding foundations from ‘Devstral’. The ‘Small’ models are designed for instruction-taking.
The new multimodal model is marketed towards developers wanting to automate coding, explore code bases and code agentic workflows. Enterprises can use Small 4 for general chat assistants, document analysis and multimodal analysis, while researchers can conduct mathematical, research and other complex reasoning tasks.
Small 4 has a total of 119bn parameters with 6bn active ones per token. It automatically switches between capabilities depending on the task, leading to a 40pc reduction in end-to-end completion time and three-times more requests per second when compared to Mistral Small 3.
