Mistral Small 4 drops 119B MoE model

Mistral Small 4 is a major architectural pivot that brings 100B+ tier reasoning to the Small family via high-density MoE. It uses 128 experts with 4 active per token to keep activated parameters at just 6.5B, maintaining high inference speed despite the 119B total size. The release unifies the Instruct, Reasoning (Magistral), and Devstral (coding) models into a single all-rounder optimized for local deployment. A new reasoning_effort parameter allows users to trade time for depth, signaling a shift toward test-time compute for local models. Apache 2.0 licensing undercuts proprietary competitors like GPT-4o mini and Claude 3.5 Haiku by offering a royalty-free alternative for 100B-class performance on consumer-accessible hardware like high-memory Macs.