huggingface.co web signal

Tsinghua-Led ScaleVLA Scales Vision-Language-Action Models to 552B on DeepSeek-V4.1 Backbone

Summary

Researchers from Tsinghua, Peking, SJTU and Galbot wire a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, running VLM scales from 2B to 552B under a shared training recipe. The 552B model needs only 10 demonstrations to beat a 2B backbone trained on 50 (87.8% vs 85.6%), reaches 45.1% on LIBERO-PRO versus 31.3% for smaller backbones, and lifts real-world charger insertion from 4/10 to 9/10 at 216ms median latency per action chunk.