Overview

Mistral-Small-3.1-24B-Instruct-2503, developed by Mistral AI, is a 24 billion parameter instruction-finetuned model that significantly enhances its predecessor, Mistral Small 3. It introduces state-of-the-art vision understanding and expands its long context capabilities up to 128k tokens without compromising text performance. The model is designed to be "knowledge-dense" and can be deployed locally, fitting within a single RTX 4090 or a 32GB RAM MacBook once quantized.

Key Capabilities

• Vision: Analyzes images and provides insights based on visual content in addition to text.
• Multilingual: Supports dozens of languages, including English, French, German, Japanese, Korean, Chinese, and Arabic.
• Agent-Centric: Offers strong agentic capabilities with native function calling and JSON outputting.
• Advanced Reasoning: Provides state-of-the-art conversational and reasoning abilities.
• Extended Context: Features a 128k context window for long document understanding.
• Apache 2.0 License: Allows for both commercial and non-commercial use and modification.

Benchmark Highlights

The model demonstrates competitive performance across various benchmarks:

• Text: Achieves 80.62% on MMLU, 88.41% on HumanEval, and 69.30% on MATH.
• Vision: Scores 64.00% on MMMU and 68.91% on Mathvista, outperforming several comparable models.
• Multilingual: Shows strong average performance at 71.18% across European, East Asian, and Middle Eastern languages.
• Long Context: Achieves 93.96% on RULER 32K, indicating robust long-context understanding.

Good For

• Fast-response conversational agents.
• Low-latency function calling.
• Local inference for hobbyists and organizations with sensitive data.
• Programming and math reasoning.
• Long document understanding and visual analysis.