AI vision models run on 67% fewer image tokens without losing accuracy

Researchers found a way to compress the image data that flows through vision-language AI models — cutting it down to a third while keeping the same accuracy. This matters because these models are expensive to run; using fewer tokens means cheaper inference, faster responses, and lower power consumption on every query.

Vision-language models are already deployed at scale in production systems; a 67% reduction in computational tokens directly lowers the cost per inference and makes real-time applications feasible on cheaper hardware.