Mistral Releases Small 4, a Unified Multimodal Reasoning Model

Mistral AI announced Small 4 on March 16, 2026, a single 119B-parameter Mixture-of-Experts model that combines instruction following, advanced reasoning, multimodal vision, and agentic coding capabilities previously spread across separate models.

Tech Insights Reporter Mar 16, 2026 4 min read Paris

TLDR

Mistral AI released Small 4, its first model to unify the capabilities of Magistral for reasoning, Pixtral for multimodal understanding, and Devstral for agentic coding into one efficient open-weight system under Apache 2.0.

Unified Architecture

Small 4 is a hybrid MoE model with 119 billion total parameters and 6 billion active parameters per token (8 billion including embeddings and output). It features a 256k context window and supports both text and image inputs.

Key innovation is the configurable reasoning_effort parameter:

"none" for fast, lightweight responses matching prior Small models.
"high" for deep step-by-step reasoning equivalent to dedicated Magistral models.

The architecture delivers 40% lower end-to-end latency in optimized setups and 3x higher throughput compared to Mistral Small 3.

Performance and Efficiency

On benchmarks like AA LCR and LiveCodeBench, Small 4 with reasoning matches or exceeds larger open models such as GPT-OSS 120B while producing significantly shorter outputs — up to 3.5-4x fewer characters in some cases.

It is designed for developers building coding agents, enterprises needing document intelligence and multimodal analysis, and researchers tackling complex reasoning tasks.

Availability and Ecosystem

The model is available immediately via Mistral API, AI Studio, and Hugging Face. It runs on NVIDIA infrastructure including NIM containers and is optimized for vLLM, SGLang, Transformers, and llama.cpp.

Mistral joined the NVIDIA Nemotron Coalition as a founding member to advance open frontier models.

Why this story matters

By consolidating specialized capabilities into one deployable model, Small 4 lowers the barrier for organizations to adopt high-performance open AI without managing multiple inference endpoints. The efficiency gains directly address cost and latency challenges in production agentic and multimodal workloads.

Sources

Mistral AI official announcement: "Introducing Mistral Small 4" (March 16, 2026) — https://mistral.ai/news/mistral-small-4/
NVIDIA NIM model card for mistral-small-4-119b-2603
Coverage in Simon Willison's blog and other technical roundups confirming release details and benchmarks

Featured Image Alt Text

Mistral AI Small 4 model announcement graphic showing unified capabilities across reasoning, vision, and coding