Présentation de Mistral Small 4

Mistral AI has announced Mistral Small 4, a new open-source model released under the Apache 2.0 license. This hybrid model unifies reasoning, multimodal (text and image), and agentic coding capabilities into a single, versatile architecture. It features a Mixture of Experts (MoE) design with 119 billion parameters, a 256k context window, and a configurable reasoning parameter to balance latency and depth of analysis. Mistral Small 4 demonstrates competitive performance on benchmarks while being more efficient by generating shorter outputs for comparable accuracy. The model is available through various platforms, including the Mistral API and Hugging Face, and is optimized for NVIDIA hardware.