Search

A vision-language-action model that runs on-device

The team distilled a 7B parameter policy down to something that fits in 8GB. Trade-offs Success rates drop roughly four points against the full model.