Mistral AI releases Mistral Small 3

Mistral AI released Mistral Small 3, a 24-billion-parameter model published under the permissive Apache 2.0 licence in both pretrained and instruction-tuned versions. The company reported 81% accuracy on MMLU and said the model matched Meta’s much larger Llama 3.3 70B on human evaluations of coding, maths and instruction-following, while running more than three times faster on the same hardware — an advantage it attributed to using fewer layers than competing models, which reduces the time taken per forward pass.

Mistral pitched the release as an open alternative to closed, proprietary small models such as OpenAI’s GPT-4o-mini, aimed at the “80%” of generative AI tasks that need solid language performance with low latency rather than frontier-scale capability: conversational assistants, function calling, and local or fine-tuned deployment.

The release came the day after DeepSeek’s R1 had renewed the argument that open-weight models trained efficiently could match proprietary systems at a fraction of the cost, adding a European entrant to a period in which open and low-cost models were closing the gap with frontier labs on several fronts at once.