Mistral Small (latest) is the Mistral Small 4 release, introduced as a unified model that brings together instruction following, reasoning, and multimodal handling in a single open-weight package. Independent reporting describes it as a 119-billion-parameter Mixture-of-Experts architecture, an unusually large configuration for the "small" family that aims to consolidate workloads previously split across separate instruct and reasoning checkpoints. The release is positioned by Mistral as a multimodal foundation, accepting text and image input while producing text output suitable for assistants, document understanding, and agent pipelines.

In practical terms, the 119B MoE design lets Mistral Small 4 activate only a subset of its parameters per request, which helps keep inference latency closer to much smaller dense models while preserving the reasoning depth of a larger backbone. Teams that previously had to route between an instruct-tuned model and a separate reasoning model can standardize on this single checkpoint, simplifying deployment for chat, structured tool use, and image-grounded tasks. Combined with its open-weight availability for local serving, it is a strong fit for organizations that want frontier-leaning multimodal capability without locking into a single proprietary stack.