ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters

arxiv.org1 min read

Researchers introduce ScaleVLA, a vision-language-action model that connects a 552-billion parameter DeepSeek-V4.1-Flash backbone to a trainable action DiT. The findings demonstrate improvements in sample efficiency, generalization, and precise manipulation while maintaining a median latency of 216 ms per action chunk. The results suggest that frontier vision-language models can serve as powerful representational foundations for robot policies. Open article →