XPeng Powers On Third Turing AI Chip and Deploys VLA+VLM Model
XPeng has announced the successful power-on of its third self-developed Turing AI chip and the deployment of its second-generation VLA+VLM cross-domain fusion model. This integration enables advanced in-car AI capabilities, including a reconstructed cockpit voice architecture that supports complex driving, parking, and navigation tasks. This milestone highlights the rapid progress of edge AI in autonomous vehicles, showcasing how custom silicon can run large multimodal models locally to improve real-time decision-making. By reducing reliance on external hardware, XPeng strengthens its vertical integration and competitiveness in the smart EV market. The Turing AI chip features 40 cores capable of running 30-billion-parameter AI models, with XPeng planning to equip future vehicles with at least three of these chips. The new VLA+VLM model supports voice-controlled roadside parking, nearby parking, and complex itinerary planning.
## BACKGROUND
A Vision-Language Model (VLM) combines computer vision with large language models to process both images and text. A Vision-Language-Action (VLA) model goes a step further by translating these multimodal inputs directly into physical actions, which is crucial for robotics and autonomous driving. Custom automotive chips like XPeng's Turing are designed to run these computationally heavy models directly on the vehicle's edge hardware without relying on cloud processing.