French startup Mistral AI has announced the release of Mistral Small 3, a 24-billion-parameter language model that the company claims matches the performance of models three times its size. According to Mistral AI, the new model achieves 81% accuracy on standard benchmarks while processing 150 tokens per second, making it comparable to Meta’s Llama 3.3 70B model while operating at significantly higher speeds.
The model is being released under the Apache 2.0 license, allowing businesses to freely modify and deploy it for their needs. Mistral AI states that the performance improvements were achieved through enhanced training optimization techniques rather than increased computing power, with the model trained on 8 trillion tokens compared to the 15 trillion typically used by comparable models. The company reports that the model was developed without reinforcement learning or synthetic training data.
According to Mistral AI, the new model is particularly suited for enterprises requiring on-premises deployment for privacy and reliability reasons, including financial services, healthcare, and manufacturing companies. The company claims the model can run on a single GPU and handle 80-90% of typical business use cases. The model is currently available through various platforms including Hugging Face, Ollama, Kaggle, Together AI, and Fireworks AI, with additional platform releases planned.
Sources: Mistral, VentureBeat