Mistral Small 4 - Mistral AI's open-source multimodal large model

What is Mistral Small 4

Mistral Small 4 is an open-source multimodal large model from Mistral AI. The model is the first to unify reasoning (Magistral), multimodal (Pixtral), and agentic coding (Devstral) capabilities within a single architecture, supports text and image input, and can be configured via the reasoning_effort parameter to flexibly switch between fast response and deep reasoning modes. The model is optimized for enterprise-grade efficiency, with 40% lower latency and 3x higher throughput, and is available on the Mistral API, Hugging Face, and NVIDIA NIM platforms.

Main features of Mistral Small 4

• Unified multi-capability architecture: The first to integrate chat instruction (Instruct), deep reasoning (Reasoning), and multimodal understanding (Multimodal) into a single model, with no need to switch between different models.
• Adjustable reasoning intensity: Flexibly controlled via the reasoning_effort parameter:
  • none: Fast response, suitable for everyday conversation.
  • high: Deep step-by-step reasoning, suitable for complex problems.
• Native multimodal processing: Supports both text and image input, enabling tasks such as document parsing, visual analysis, and image-text understanding.
• Agentic coding capability: Supports development scenarios such as code generation, codebase exploration, and automated programming workflows.
• Long context processing: Supports a 256K context window, long document analysis, and long conversations
• Enterprise-grade efficiency: Compared with the previous generation, latency is reduced by 40% and throughput is increased by 3x, supporting efficient deployment.

Key information and usage requirements for Mistral Small 4

• Architecture：Mixture of Experts (MoE)
• Number of experts: 128 experts, with 4 activated per token
• Total parameter count: 119 billion (119B)
• Activated parameter count: 6 billion/token (8 billion including the embedding layer)
• Context window：256K tokens
• Open-source license：Apache 2.0
• Hardware requirements
  • Minimum configuration: 4× NVIDIA HGX H100 or 2× HGX H200 or 1× DGX B200
  • Recommended configuration: 4× NVIDIA HGX H100 or 4× HGX H200 or 2× DGX B200

Core advantages of Mistral Small 4

• Integrated unification: The first to unify the three major capabilities of reasoning, multimodality, and Agent coding into a single model, with no need to switch between multiple models.
• Flexible reasoning: Freely switch between fast response and deep thinking modes via the reasoning_effort parameter, allocating compute as needed.
• Extreme efficiency: At comparable performance, output length is significantly shorter, directly reducing inference costs and improving user experience.
• Truly open source: The Apache 2.0 license supports commercial use and deep customization, and works with NVIDIA NeMo for domain fine-tuning.
• Ecosystem integration: As a founding member of the NVIDIA Nemotron Alliance, it receives full-stack optimization support from hardware to deployment tools.
• Enterprise value: Lower token costs and more stable quality make large-scale AI deployment more economically viable.
• Technical value: High "performance per token" simplifies model selection and reduces fine-tuning iterations and reliance on backup systems.

How to use Mistral Small 4

• Via the official Mistral platform: Call it directly in the Mistral API or AI Studio, with no need to build your own infrastructure, making it suitable for quick onboarding and prototype validation.
• Via Hugging Face: Download the model weights from the Hugging Face repository and use open-source frameworks such as Transformers, vLLM, llama.cpp, and SGLang for local deployment and inference.
• Via the NVIDIA platform: Test the model's performance for free on build.nvidia.com, or use NVIDIA NIM for production-grade containerized deployment to get out-of-the-box optimized inference performance.
• Via fine-tuning customization: Use the NVIDIA NeMo framework to fine-tune the model for domain-specific needs and create a customized version that meets specific business requirements.
• Configure reasoning effort: Control behavior through the reasoning_effort parameter when calling: set it to "none" for fast responses, or set it to "high" to activate deep reasoning mode.
• Hardware requirements: Local deployment requires at least 4x HGX H100 or 1x DGX B200 level compute power, and doubling the recommended configuration ensures optimal performance.

Mistral Small 4 project address

• Project official website：https://mistral.ai/news/mistral-small-4
• HuggingFace model library：https://huggingface.co/collections/mistralai/mistral-small-4

Comparison of Mistral Small 4 with similar competitors

Application scenarios of Mistral Small 4

• Intelligent programming: The model can automatically generate code, fix bugs, and understand the architecture of large codebases, improving development efficiency.
• Enterprise customer service: Handle routine inquiries and complex complaints through adjustable reasoning modes, reducing the cost of manual intervention.
• Document analysis: The model can parse long documents, contracts, and cross-file related information, supporting deep processing with a 256K context.
• Visual understanding: Supports recognizing invoices, charts, and image content, enabling intelligent information extraction that combines text and images.
• Scientific research assistance: The model can complete mathematical derivations, paper interpretation, and experimental design, providing academic support with step-by-step reasoning.