Mistral Small 3

Mistral Small 3 is the most efficient and versatile model of Mistral. Pre-trained and instructed version, Apache 2.0, 24B, 81% MMLU, 150 token/s. No synthetic data so great base for anything reasoning.

Mistral Small 3 is a high-performance, open-source 24-billion-parameter model designed for efficiency and versatility. Released under the Apache 2.0 license, it excels in low-latency tasks, delivering 150 tokens per second with 81% accuracy on MMLU benchmarks. It rivals larger models like Llama 3.3 70B and proprietary alternatives such as GPT4o-mini, offering competitive performance while being over three times faster on the same hardware. Ideal for conversational AI, rapid function execution, and domain-specific fine-tuning, Mistral Small 3 is optimized for local deployment, making it accessible for hobbyists and organizations handling sensitive data. Available on platforms like Hugging Face, Ollama, and Kaggle, it supports diverse use cases, from fraud detection to healthcare triaging. Its lightweight architecture and lack of synthetic data make it a robust foundation for building advanced reasoning capabilities, fostering innovation within the open-source community.