Mistral Small 3

Mistral Small 3 is a latency-optimized 24-billion-parameter language model from Mistral AI, released on 30 January 2025 under the Apache 2.0 license (both the base and instruction-tuned versions). It was designed for fast conversational responses and low-latency function calling in task automation.

The model achieves 81% on the MMLU benchmark and generates about 150 tokens per second. According to Mistral AI, it matches the quality of the much larger Llama 3.3 70B while being more than 3× faster on the same hardware, and it is competitive with GPT-4o mini and Qwen 32B.

Thanks to its compact size, Mistral Small 3 can run locally — for example on a single NVIDIA RTX 4090 or a MacBook with 32 GB of memory — making it a good choice for local deployments, domain-specific fine-tuning and privacy-sensitive use cases.

The model is available with open weights (base and instruct versions) on Hugging Face, via the Mistral API (la Plateforme; mistral-small-latest / mistral-small-2501), and through partners: Ollama, Kaggle, Together AI, Fireworks AI and IBM watsonx (with announced support for NVIDIA NIM, Amazon SageMaker, Groq, Databricks and Snowflake, among others).