Mistral Small 4 Unifies Reasoning, Multimodal, and Coding in One Open-Source Model

Mistral Small 4, the latest release from French AI company Mistral AI, combines reasoning, multimodal processing, and agentic coding capabilities into a single open-source model. The company said the model is available under the Apache 2.0 license, making it freely usable, modifiable, and distributable.

What Is Mistral Small 4?

Mistral Small 4 consolidates the capabilities of three previous specialized models: Magistral for reasoning, Pixtral for multimodal tasks, and Devstral for agentic coding. Previously, users had to select separate models depending on their task. The new release eliminates that requirement by delivering all three functions within one architecture.

Mistral Small 4 Architecture and Key Specifications

The model uses a Mixture of Experts (MoE) architecture with 128 experts, activating 4 per token. It carries 119 billion total parameters, with 6 billion active per token. Moreover, the model supports a 256,000-token context window, enabling long-form document analysis and extended interactions.

A configurable reasoning_effort parameter allows users to toggle between fast, low-latency responses and deep, step-by-step reasoning. Setting the parameter to none produces lightweight responses comparable to Mistral Small 3.2, while setting it to high delivers detailed reasoning comparable to the Magistral model series.

Performance Benchmarks

According to Mistral AI, the model achieves a 40% reduction in end-to-end completion time in latency-optimized configurations. Furthermore, it delivers three times more requests per second compared to Mistral Small 3 in throughput-optimized setups.

On the AA LCR benchmark, Mistral Small 4 scores 0.72 using just 1,600 characters of output. In comparison, Qwen models require 5,800 to 6,100 characters for comparable performance. On LiveCodeBench, the model outperforms GPT-OSS 120B while producing 20% less output. Shorter outputs translate directly to lower latency and reduced artificial intelligence inference costs.

Infrastructure Requirements

The minimum hardware configuration requires 4x NVIDIA HGX H100, 2x NVIDIA HGX H200, or 1x NVIDIA DGX B200. The recommended setup for optimal performance is 4x NVIDIA HGX H100, 4x NVIDIA HGX H200, or 2x NVIDIA DGX B200.

Mistral AI said it collaborated closely with NVIDIA to optimize inference for both vLLM and SGLang frameworks. The company also stated it is joining the NVIDIA Nemotron Coalition as a founding member. Additionally, the model is available on Hugging Face, llama.cpp, Transformers, and other platforms.

Intended Use Cases and Availability

Mistral Small 4 targets three primary user groups. Developers can use it for coding automation, codebase exploration, and agentic workflows. Enterprises can deploy it for general chat assistants, document understanding, and multimodal analysis. Researchers can apply it to mathematics, research tasks, and complex reasoning problems.

The model is available through the Mistral API, AI Studio, and the Hugging Face repository. Developers can prototype the model for free on NVIDIA accelerated computing infrastructure. For production deployment, Mistral Small 4 is available as an NVIDIA NIM, providing containerized inference. Domain-specific fine-tuning is supported through NVIDIA NeMo. Enterprise customers seeking custom fine-tuning or on-premises deployment can contact Mistral AI directly.

“By unifying instruct, reasoning, and multimodal capabilities, Mistral Small 4 simplifies AI integration and empowers users to tackle a wider range of tasks with a single, adaptable tool.”

Mistral AI

The open-source release continues Mistral AI’s approach of providing AI models under permissive licenses. The apps and software ecosystem around the model already includes support from major inference frameworks, lowering the barrier for teams looking to deploy or fine-tune the model at scale.