XPeng CEO Sets Multiple OKRs for VLA Gen2

🔑 Enhanced Key Takeaways

• •XPeng launched VLA 2.0 on March 2, 2026, during a media event alongside the 2026 X9 MPV, with initial deployment to Ultra versions of P7, G7, and X9 in late March.[1][5]
• •VLA 2.0 features a novel 'Vision-Implicit Token-Action' architecture, eliminating language translation for end-to-end visual-to-action generation, and runs billions of parameters on 2250 TOPS hardware.[4]
• •Volkswagen is the first external customer for VLA 2.0, with XPeng planning global rollout in 2027; Robotaxi public road testing has begun using this model.[2][4][5]

🛠️ Technical Deep Dive

• •Adopts 'Vision-Implicit Token-Action' (VIA) path, bypassing traditional vision-language-action pipeline for direct end-to-end generation from visual signals to actions.[4]
• •Optimized via full-cycle 'chip-operator-model' integration, enabling billion-parameter scale on vehicle-end hardware with 2250 TOPS compute, vs. industry tens of millions.[4]
• •Inference speed 12x faster than prior generation; 23% improved driving efficiency, surpassing L2 systems and some Robotaxi models in urban rush hour tests.[5][8]
• •Supports Level 4-oriented driving with multimodal physical world modeling, applicable to autonomous driving and robotics.[3]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline