Overview

Mistral-Small-3.2-24B-Instruct-2506 is an updated 24 billion parameter instruction-tuned model from Mistral AI, enhancing its predecessor, Mistral-Small-3.1-24B-Instruct-2503. This iteration focuses on refining core capabilities crucial for reliable AI applications.

Key Improvements & Capabilities

Enhanced Instruction Following: Demonstrates better adherence to precise instructions, with significant gains on benchmarks like Wildbench v2 (65.33%) and Arena Hard v2 (43.1%).
Reduced Repetition Errors: Significantly decreases infinite generations and repetitive outputs, showing a 2x reduction on challenging prompts.
Robust Function Calling: Features an improved and more reliable function calling template, facilitating better integration with tools and agents.
Multimodal Support: Retains vision capabilities, allowing for reasoning over image inputs, as demonstrated in examples involving visual scenarios.
General Performance: Maintains or slightly improves performance across various categories, including STEM benchmarks like MMLU Pro (69.06%), MBPP Plus - Pass@5 (78.33%), and HumanEval Plus - Pass@5 (92.90%).

Recommended Usage

This model is well-suited for applications requiring precise instruction adherence, reliable function calling, and multimodal understanding. It is recommended to use a low temperature (e.g., 0.15) and provide a system prompt for optimal performance. The model can be deployed efficiently using vLLM (recommended) or Transformers frameworks.