Mistral Small

Mistral Small is a latency-optimized, 24B-parameter language model: it delivers high-end performance, including multimodal and multilingual capabilities, while maintaining the efficiency and speed required for production-grade, low-latency applications.

Mistral Small is a powerful, efficiency-focused model (24B parameters) that sets a new performance standard in its weight class: it consistently outperforms comparable models like GPT-4o Mini and Gemma 3 on key benchmarks. The latest version, Mistral Small 3.1, offers a massive 128k context window, multimodal understanding, and multilingual support. It's built for speed, delivering inference at up to 150 tokens per second, making it ideal for fast conversational agents, function calling, and local deployment on hardware like a single NVIDIA RTX 4090.

https://docs.mistral.ai/getting-started/models/models_overview/