Mistral Releases Small 4 Unified Open-Source Model

Four capabilities, one model

Paris-based Mistral AI released Mistral Small 4 on March 16, consolidating capabilities from four previously separate product lines into a single 119-billion-parameter model. The unified system combines instruction-following from Mistral Small, extended reasoning from Magistral, image understanding from Pixtral, and agentic coding from Devstral.

The model uses a mixture-of-experts architecture with 128 experts but activates only 6 billion parameters per token, keeping inference costs closer to a much smaller dense model. Released under the permissive Apache 2.0 license, the model allows commercial use, fine-tuning, and redistribution without royalty requirements.

Configurable reasoning effort

A notable feature allows developers to adjust reasoning depth per request through a reasoning_effort parameter. Setting it to 'none' produces fast responses equivalent to the previous Mistral Small 3.2, while 'high' enables step-by-step reasoning with verbosity comparable to earlier Magistral models.

This design avoids forcing teams to route between different models depending on task complexity. Instead, a single endpoint handles both lightweight queries and complex analysis, with teams paying for deeper reasoning only when needed.

Efficiency claims and deployment

Mistral reports that Small 4 delivers 40 percent lower end-to-end completion time and three times the throughput compared to Small 3 in optimized configurations. The company emphasizes that shorter outputs at equivalent accuracy reduce latency, inference costs, and downstream processing overhead.

The model supports a 256,000-token context window and is compatible with common open inference frameworks including vLLM, llama.cpp, SGLang, and Transformers. Minimum hardware requirements include four NVIDIA HGX H100 units, two HGX H200 units, or one DGX B200 system.