XPeng 2nd-Gen VLA Rolls Out This Month

🔑 Enhanced Key Takeaways

• •XPeng VLA 2.0 adopts a 'Vision-Implicit Token-Action' architecture, bypassing traditional language translation for direct visual-to-action generation.[1][2][3]
• •The model was trained on nearly 100 million unannotated driving video clips, equivalent to 65,000 years of human driving experience.[3][4][6]
• •Powered by Turing AI chips delivering 2,250 TOPS per chip, enabling 12x inference efficiency gains and deployment across cars, robots, and flying cars.[1][2][6]
• •VLA 2.0 serves as XPeng's first mass-produced physical world model, capable of self-evolving learning and cross-domain applications.[3][5]

🛠️ Technical Deep Dive

• •Architecture shift to 'Vision-Implicit Token-Action' (V-Implicit Token-A) eliminates V-L-A language bottleneck, enabling end-to-end V-A output for human-like reflexes.[1][2][3]
• •Trained without data annotation using ~100 million real-world driving clips covering long-tail scenarios; generates adversarial scenarios for self-improvement.[3][4][6]
• •Deployed via custom compiler on Turing AI chip (2,250 TOPS/chip); model parameters 10x larger than mainstream, with 12x inference efficiency.[1][6]
• •Combines with VLT (Vision-Language-Thinking) and VLM for high-order capabilities like conversation, walking, and interaction in robots like IRON.[2][3][6]
• •XPeng World Base Models trained at 1B, 3B, 7B, 72B parameters; four-stage factory process: pre-training, distillation, continued training, vehicle deployment.[7]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline