Mistral Small 4 (mistral.ai/news/mistral-small-4/) is an open-source, hybrid foundation model developed by Mistral AI. It unifies the capabilities of Mistral’s specialized flagships—Magistral (deep reasoning), Pixtral (multimodal vision understanding), and Devstral (agentic coding)—into a single architecture. Built on a 119B-parameter sparse Mixture-of-Experts (MoE) structure activating 6B parameters per token across 128 experts (4 active), it features configurable reasoning effort (reasoning_effort), a 256K context window, 40% lower completion latency, and day-0 NVIDIA NIM deployment.

Previously, developers and enterprise teams had to route tasks across disparate models—switching between fast instruction models for chat, specialized reasoning engines for logic/math, and vision models for image analysis. Mistral Small 4 eliminates this fragmentation by combining all three domains into one sparse, high-efficiency architecture.

Architectural Profile & Specifications at a Glance

Feature / Dimension	Mistral Small 4 Technical Specification	Production & Developer Advantage
Total / Active Parameters	119B Total / 6B Active (8B w/ embeddings)	High knowledge capacity with low inference latency and active compute costs.
MoE Architecture	128 Total Experts (4 Active per Token)	Granular specialization and high parameter efficiency during routing.
Supported Modalities	Native Multimodal (Text and Image Inputs)	Ingests documents, charts, visual queries, and text codebases concurrently.
Context Window	256,000 Tokens (256K)	Ingests entire repositories, multi-page PDFs, and long conversation histories.
Reasoning Control	Dynamic reasoning_effort parameter	Switch from instant zero-overhead responses (none) to deep CoT (high).
Throughput & Speed	40% lower completion time / 3x throughput vs. Small 3	Significant cost reduction and faster end-to-end token generation.
Licensing & Ecosystem	Fully Open Source / NVIDIA Nemotron Coalition	Permissive fine-tuning, self-hosting, and day-0 NVIDIA NIM container support.

Key Capabilities and Innovations

• Unified Flagship Intelligence: Consolidates three distinct model lineages into a single checkpoint: general instruction following (Mistral Small), extended step-by-step thinking (Magistral), visual comprehension (Pixtral), and autonomous coding workflows (Devstral).

• Dynamic Reasoning Effort Adjustment: Allows developers to configure runtime thinking intensity via the API:

  • reasoning_effort="none": Fast, concise responses for standard chat and low-latency workflows.

  • reasoning_effort="high": Activates deep, step-by-step reasoning traces for complex STEM problems, logic proofs, and codebase refactoring.

• Concise, High-Density Reasoning: Generates significantly shorter reasoning traces than competing open-weight models while matching or surpassing frontier benchmarks, reducing token consumption, latency, and overall API/compute expenses.

• Enterprise & Agent-Ready Serving: Available for instant prototyping on build.nvidia.com and production deployment via NVIDIA NIM, with native support for FP4/NVFP4 quantization, Eagle speculative decoding heads, vLLM, and SGLang.