119B / 6.5B active (MoE)
Apache 2.0
Open-weight multimodal MoE model positioned as a unified successor across Mistral's Instruct, Reasoning (Magistral), and Devstral usage patterns in one checkpoint family. Official release materials describe Mistral Small 4 as the first Small-family model to combine fast instruct behavior, on-demand deep reasoning, and agentic coding capabilities in a single system, alongside multimodal text+image input support. Architecture details cite 128 experts with 4 active experts per token, 119B total parameters, and approximately 6B-6.5B activated parameters per token (with release notes also describing around 8B active including embedding/output layers), plus a 256K context window for long-form tasks. The model supports configurable reasoning via per-request `reasoning_effort` controls, allowing fast low-latency operation (`none`) or deeper stepwise reasoning (`high`) depending on task complexity. Official positioning emphasizes strong function-calling and structured-output workflows, multilingual operation, and broad applicability across chat, coding, tool-use, and document/image analysis scenarios. Release materials also provide efficiency-oriented companion checkpoints: an NVFP4 quantized variant for lower-cost inference and an Eagle speculative-decoding head for faster generation. Announcement and card claims include improved serving efficiency relative to prior Small-family baselines, including reported 40% reduction in end-to-end completion time and 3x higher request throughput in optimized setups.