Mistral Small 3

Mistral Small 3 is a highly efficient and versatile 24B-parameter model optimized for low latency, making it competitive with much larger models like Llama 3.3 70B and Qwen 32B. Released under the Apache 2.0 license, it is designed for robust language and instruction-following tasks, achieving over 81% accuracy on MMLU benchmarks and processing 150 tokens per second. Unlike some models, Mistral Small 3 is not trained with reinforcement learning or synthetic data, providing a strong base for reasoning tasks.

This model is particularly well-suited for scenarios requiring fast-response conversational assistance, low-latency function calling, and fine-tuning for specific domains. It is ideal for local deployment, capable of running on hardware like a single RTX 4090 or a Macbook with 32GB RAM. Use cases span across industries such as financial services for fraud detection, healthcare for customer triaging, and manufacturing for on-device command and control.

Mistral Small 3 is available on multiple platforms including Hugging Face, Ollama, Kaggle, Together AI, and Fireworks AI, with upcoming support on NVIDIA NIM, Amazon SageMaker, Groq, Databricks, and Snowflake. The model supports open-source customization and deployment, encouraging the community to enhance and build upon its capabilities.