Qwen Announces VLA Model That Controls Robots Across Different Hardware
Alibaba’s Qwen team has unveiled Qwen-VLA, a vision-language-action model capable of controlling robots of varying designs without per-platform fine-tuning. Built on the Qwen3.5-4B vision-language backbone with a 1.15B-parameter action decoder, the model unifies three task types — manipulation, navigation, and trajectory prediction — and switches between different robot bodies by simply changing the text instruction.
Qwen-VLA matches or surpasses specialized systems trained for individual tasks, scoring 97.9% on LIBERO, 87.2% on RoboTwin-Hard, and 83.6% / 76.9% success rates on the ALOHA dual-arm platform in familiar and unfamiliar settings respectively. A technical report and GitHub repository are now available, though the model weights themselves have not yet been released.