Mistral unveils Small 3, a fast, 24B-parameter AI model to rival GPT4o-mini & other models

Mistral has announced the release of its new large language model, Mistral Small 3, a 24-billion-parameter model optimized for low latency, and released under the Apache 2.0 license. This model is positioned as a competitive alternative to larger models like Llama 3.3 70B and Qwen 32B, and as an open replacement for proprietary models such as GPT4o-mini.

Mistral Small 3 matches the performance of Llama 3.3 70B instruct while achieving more than three times the speed on equivalent hardware. It is designed to handle 80% of generative AI tasks, focusing on robust language and instruction-following capabilities with minimal latency. The model's architecture is optimized for local deployment, featuring fewer layers than its competitors, which reduces the time required per forward pass.

Mistral has released both pre-trained and instruction-tuned checkpoints, aiming to provide a strong foundation for further development. Unlike some models, Mistral Small 3 is not trained with reinforcement learning or synthetic data, positioning it earlier in the production pipeline. This makes it a suitable base for developing reasoning capabilities in future models.