XPeng to Release Second-Gen VLA Lite Distilled Model for Single Turing Chip Vehicles
XPeng Motors announced that its second-generation VLA Lite distilled model will be rolled out in September, starting with the G9L Max version. This update targets Max vehicles equipped with a single Turing chip, enabling L2 to L4 autonomous driving capabilities. Deploying a distilled Vision-Language-Action (VLA) model on single-chip edge hardware demonstrates practical progress in running advanced autonomous driving AI on resource-constrained platforms. This approach could accelerate the adoption of high-level autonomous driving features without requiring expensive multi-chip hardware upgrades. XPeng achieved this deployment using learnable token compression and distillation training, which reduces the number of active tokens for high-quality visual understanding without losing input information or model capacity. As the base model improves, the distilled Turing VLA 2.0 Lite model also receives performance upgrades.
## BACKGROUND
A Vision-Language-Action (VLA) model is a multimodal AI that processes visual inputs and language instructions to output physical actions, which is crucial for robotics and autonomous driving. Knowledge distillation is a machine learning technique that transfers knowledge from a large, complex model to a smaller, faster one so it can run on lower-power hardware. Token compression further reduces computational costs by pruning or merging redundant tokens in large models.