awesome-ai-radar
← Back to Radar

Mistral Small 4 Unifies Instruct, Reasoning, and Multimodal in One Open-Source MoE

open-source self-hosting multimodal agents

What happened

Mistral AI released Mistral Small 4 on March 16, 2026 — a 119-billion-parameter Mixture-of-Experts model with 128 experts and a 256k context window, licensed under Apache 2.0. The model unifies instruction following, configurable reasoning, multimodal understanding (text + images), and agentic coding into a single deployment. Compared to Small 3, it delivers 40% lower latency and 3x higher throughput. Available via Mistral API, Hugging Face, and NVIDIA NIM containers with vLLM and llama.cpp support.

Why it matters

Small 4 eliminates the need to run separate models for different tasks — one deployment handles chat, reasoning, vision, and tool use. The Apache 2.0 license and broad inference framework support make it a strong candidate for teams that want frontier-adjacent capabilities without vendor lock-in. The configurable reasoning toggle lets developers trade latency for depth per request.

Who should pay attention

  • Teams self-hosting open-source models who want to consolidate multiple deployments into one
  • Developers building agentic applications that need reasoning, vision, and tool use in a single model
  • Organizations in regulated industries that require on-premises deployment with an open license