CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model

CoTinyVLA uses a 0.9B-parameter action model built on Qwen3.5-0.8B, combining dual-view temporal inputs, hierarchical chain-of-thought distillation from a 35B teacher, and paraphrase augmentation. On 10,030 perturbed LIBERO-Plus tasks, it reports 90.8% Spatial, 87.3% Object, 86.6% Goal, and 80.7% Long success, exceeding the strongest 7B baseline by 2.8 to 15.9 points. Closed-loop inference peaks at 2.25 GiB of allocated GPU memory.