Mistral Small 4 Released

Mistral Small 4 has been released as a hybrid model that combines Instruct, Reasoning(Magistral), and Devstral capabilities into one.

• MoE 128 experts / 4 active, 119B total parameters, 6.5B activated per token
• Supports 256k context and text+image input, generating text output
• Can switch between instant response and reasoning mode via per-request reasoning_effort, and also supports system prompts, function calling, and JSON output
• Recommends reasoning_effort='high' for complex tasks and none mode for lightweight tasks, along with temperature guidance
• Supports multiple languages including English, French, Spanish, German, Chinese, Japanese, and Korean, and is distributed under Apache 2.0

According to the performance description, in the latency-optimized setting, end-to-end completion time was reduced by 40%, and in the throughput-optimized setting, requests processed per second increased 3x compared to Mistral Small 3.

Derivative checkpoints for further efficiency were also presented.

• mistralai/Mistral-Small-4-119B-2603-eagle: for speculative decoding
• mistralai/Mistral-Small-4-119B-2603-NVFP4: 4-bit NVFP4 quantized checkpoint

In benchmark comparisons, the reasoning mode matched or outperformed GPT-OSS 120B, achieving AA LCR 0.72 with 1.6K characters of output. To achieve comparable performance, the Qwen series required 5.8K~6.1K output, and on LiveCodeBench it produced higher results than GPT-OSS 120B while using 20% less output.

Deployment is supported via vLLM, llama.cpp, LM Studio, SGLang, and transformers, and it can be used for document understanding, image analysis, coding agents, research assistance, and fine-tuning.