Mistral Small 4 release notes / what's new

Mistral AI has announced the release of Mistral Small 4, a unified model that consolidates the capabilities of Magistral (reasoning), Pixtral (multimodal), and Devstral (agentic coding) into a single architecture. This release eliminates the need for users to switch between specialized models for different tasks, providing a versatile tool with configurable reasoning effort and high efficiency.

Unified Model Capabilities

Mistral Small 4 integrates three previously distinct model strengths into one system:

  • Reasoning: Capabilities derived from Magistral.
  • Multimodal: Image and text input support derived from Pixtral.
  • Agentic Coding: Coding automation and agentic workflows derived from Devstral.

This unification allows the model to serve as a chat assistant, research partner, or coding agent without requiring the user to change models based on the task.

Technical Architecture and Specifications

Mistral Small 4 is a hybrid model utilizing a Mixture of Experts (MoE) architecture. Key technical specifications include:

  • Parameter Count: 119B total parameters, with 6B active parameters per token (8B including embedding and output layers).
  • Expert Configuration: 128 experts, with 4 active per token.
  • Context Window: 256k tokens, supporting long-form document analysis and interactions.
  • Input Support: Native multimodality allowing for both text and image inputs.
  • Licensing: Released under the Apache 2.0 license.

Configurable Reasoning Effort

The model introduces a reasoning_effort parameter that allows users to dynamically adjust the depth of the model's processing:

  • reasoning_effort="none": Provides fast, lightweight responses for everyday tasks, mirroring the chat style of Mistral Small 3.2.
  • reasoning_effort="high": Enables deep, step-by-step reasoning for complex problems, matching the verbosity of previous Magistral models.

Performance and Efficiency

Mistral Small 4 demonstrates significant improvements in throughput and latency compared to Mistral Small 3:

  • Latency: 40% reduction in end-to-end completion time in latency-optimized setups.
  • Throughput: 3x increase in requests per second in throughput-optimized setups.

In benchmark comparisons, Mistral Small 4 with reasoning matches or surpasses GPT-OSS 120B across three benchmarks while generating shorter outputs. On the AA LCR benchmark, it scores 0.72 using 1.6K characters, whereas Qwen models require 3.5-4x more output (5.8-6.1K characters) for similar performance. On LiveCodeBench, it outperforms GPT-OSS 120B while producing 20% less output.

Deployment and Infrastructure Requirements

Mistral Small 4 is optimized for NVIDIA hardware and is a founding member of the NVIDIA Nemotron Coalition. Deployment options include:

Infrastructure Tiers

  • Minimum Infrastructure: 4x NVIDIA HGX H100, 2x NVIDIA HGX H200, or 1x NVIDIA DGX B200.
  • Recommended Setup: 4x NVIDIA HGX H100, 4x NVIDIA HGX H200, or 2x NVIDIA DGX B200 for optimal performance.

Availability and Integration

  • Platforms: Available via Mistral API, AI Studio, and Hugging Face.
  • Frameworks: Supported on vLLM, llama.cpp, SGLang, and Transformers.
  • NVIDIA Integration: Available as an NVIDIA NIM for containerized inference and customizable via NVIDIA NeMo for domain-specific fine-tuning.
  • Prototyping: Free prototyping is available at build.nvidia.com.

Sources

Related

  • Dispatch
    Mistral Small 3.1 Release Notes

    Mistral AI has released Mistral Small 3.1, an Apache 2.0 licensed model featuring multimodal understanding, a 128k token context window, and performance that surpasses GPT-4o Mini and Gemma 3.

  • Dispatch
    Mistral AI Devstral 2 and Mistral Vibe CLI Release

    Mistral AI has released Devstral 2 (123B) and Devstral Small 2 (24B), open-weight coding models that achieve state-of-the-art performance on SWE-bench Verified with high cost-efficiency, alongside the open-source Mistral Vibe CLI for autonomous code automation.

  • Dispatch
    Mistral Small 3 Release Notes

    Mistral AI has released Mistral Small 3, a latency-optimized 24B-parameter model under the Apache 2.0 license that rivals the performance of models three times its size.

  • Dispatch
    Mistral Large and Mistral Small Release

    Mistral AI has released Mistral Large, a flagship model with top-tier reasoning and multilingual capabilities, and Mistral Small, a model optimized for low-latency workloads.

  • Dispatch
    Mistral AI Devstral Models Release

    Mistral AI has released Devstral Medium and Devstral Small 1.1, providing state-of-the-art agentic coding capabilities with a focus on generalization across different prompts and agentic scaffolds.