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. It is the first model to unify inference (magistral), multimodal (pixtral), and agent coding (Devstral) capabilities into a single architecture. It supports text and image input and can be used... reasoning_effort Flexible parameter switching between fast response and deep inference modes. The model is optimized for enterprise-level efficiency, reducing latency by 40% and increasing throughput by 3x. It is available on Mistral API, Hugging Face, and NVIDIA NIM platform.

Main functions of Mistral Small 4

• Unified multi-capability architectureFor the first time, chat instructions, deep reasoning, and multimodal understanding are integrated into a single model, eliminating the need to switch between different models.
• Adjustable reasoning strength:pass reasoning_effort Flexible parameter control:
  • noneQuick response, suitable for everyday conversations.
  • highDeep stepwise reasoning, suitable for complex problems.
• Native multimodal processingIt supports both text and image input, enabling tasks such as document parsing, visual analysis, and image-text understanding.
• Intelligent agent coding abilityIt supports development scenarios such as code generation, code library exploration, and automated programming workflows.
• Long context processingSupports 256K context windows, long document analysis, and long conversations.
• Enterprise-level efficiencyCompared to its predecessor, latency is reduced by 40%, throughput is increased by 3 times, and efficient deployment is supported.

Key information and usage requirements for Mistral Small 4

• ArchitectureMixture of Experts (MoE)
• Number of experts128 experts, 4 activated per token
• Total number of parameters119 billion (119B)
• Activation parameter quantity6 billion per token (8 billion including the embedded layer)
• Context window256K tokens
• Open source licenseApache 2.0
• Hardware Requirements
  • Minimum configuration: 4 × NVIDIA HGX H100 or 2 × HGX H200 or 1 × DGX B200
  • Recommended configuration4× NVIDIA HGX H100 or 4× HGX H200 or 2× DGX B200

Mistral Small 4's core advantages

• IntegratedFor the first time, it unifies the three major capabilities of reasoning, multimodal, and agent programming into a single model, eliminating the need to switch between multiple models.
• Flexible reasoningThe reasoning_effort parameter allows you to freely switch between fast response and deep thinking modes and allocate computing power as needed.
• Ultimate efficiencyWith the same performance, the output length is significantly shorter, which directly reduces inference costs and improves user experience.
• Truly open sourceThe Apache 2.0 license supports commercial use and deep customization, allowing for fine-tuning in conjunction with NVIDIA NeMo.
• Ecosystem bindingAs a founding member of the NVIDIA Nemotron Alliance, we receive full-stack optimization support, from hardware to deployment tools.
• Enterprise ValueLower token costs and more stable quality make large-scale AI deployments more economically feasible.
• Technological valueHigh performance per token simplifies model selection, reduces fine-tuning iterations, and reduces reliance on backup systems.

How to use Mistral Small 4

• Through the official Mistral platformIt can be called directly in the Mistral API or AI Studio without the need to build your own infrastructure, making it suitable for quick start-up and prototype verification.
• via Hugging FaceDownload 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.
• Through NVIDIA platformTest your model for free at build.nvidia.com, or deploy it in a production-ready containerized manner using NVIDIA NIM for optimized inference performance out of the box.
• Through fine-tuning and customizationUse the NVIDIA NeMo framework to fine-tune the model for specific business needs, creating a customized version.
• Configure inference strengthThe behavior is controlled by the reasoning_effort parameter when the function is invoked: setting it to "none" enables a fast response, while setting it to "high" activates the deep reasoning mode.
• Hardware RequirementsLocal deployment requires a minimum of 4x HGX H100 or 1x DGX B200 level computing power; it is recommended to double the configuration to ensure optimal performance.

Mistral Small 4 project address

• Project official websitehttps://mistral.ai/news/mistral-small-4
• HuggingFace model libraryhttps://huggingface.co/collections/mistralai/mistral-small-4

Comparison of Mistral Small 4 with similar competing products

Application scenarios of Mistral Small 4

• Intelligent ProgrammingThe 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 models to reduce the cost of manual intervention.
• Document AnalysisThe model can parse long documents, contracts, and cross-file association information, and supports 256K context depth processing.
• Visual understandingIt supports the recognition of invoices, charts, and images, enabling intelligent information extraction that combines text and images.
• Research supportThe model can perform mathematical derivation, paper interpretation, and experimental design, providing academic support for step-by-step reasoning.