Mistral's new open model fuses reasoning, vision, and coding

Mistral Small 4 is a unified open-source model that combines deep reasoning, multimodal vision, and agentic coding into a single package. Instead of routing prompts across Magistral for reasoning, Pixtral for image input, and Devstral for coding tasks, developers can now run everything through one Apache 2.0 licensed model with a massive 256k context window.

How it works under the hood

The engine behind Small 4 is a sparsely activated Mixture of Experts (MoE) architecture. While the model contains 119B total parameters across 128 experts, it only routes each token through 4 experts at runtime. That means just 6B parameters are active per token, or 8B including the embedding and output layers.

It also introduces a flexible reasoning_effort switch. Set it to none, and it acts like a lightweight instruct engine matching the conversational speed of Mistral Small 3.2. Dial it up to high, and it engages verbose, step-by-step thinking equivalent to Magistral's dedicated reasoning models.

Benchmarked against Mistral Small 3, Small 4 delivers 40% lower latency in latency-optimized setups and pushes 3x more requests per second in throughput-oriented environments.

More importantly, it doesn't waste tokens. On the AA LCR benchmark, Small 4 scored 0.72 using only 1.6K characters, whereas comparable Qwen models generated 5.8K to 6.1K characters—nearly 4x the output—to reach similar accuracy. On LiveCodeBench, it beat GPT-OSS 120B while cutting generated output volume by 20%.

Running the model and hardware specs

Because it carries 119B total parameters, self-hosting requires serious enterprise silicon. The minimum deployment footprint is 4x NVIDIA HGX H100, 2x HGX H200, or a single DGX B200. For production workloads, Mistral recommends 4x H100, 4x H200, or 2x B200.

Day-0 integration is already live across vLLM, llama.cpp, SGLang, and Hugging Face Transformers. Developers who want to test it right away can use the Mistral API and AI Studio, run free prototypes at build.nvidia.com, or deploy it as a pre-packaged NVIDIA NIM container.

Why it matters

Small 4 eliminates the headache of orchestrating separate specialized models for vision, code generation, and logical reasoning. For teams building autonomous coding agents or enterprise document assistants, having a single adaptable model with dialable reasoning effort drastically simplifies backend architecture.

In our view, Mistral's obsession with token efficiency is the standout achievement here. By delivering frontier-level scores without producing bloated chains of thought, Small 4 directly lowers token consumption, reduces per-query latency, and cuts cloud inference bills where it counts.