XPENG has commenced mass production of its in-house VLA2.0 foundation model, which achieved over 50% assisted driving mileage share within its first month of deployment. The company also reported a 1,010% increase in per-GPU training efficiency and is accelerating rollout of Robotaxi and IRON humanoid robots.
XPENG has moved its in-house foundation model VLA2.0 into full-scale mass production following technical validation, achieving over 50% share of assisted driving mileage in its first month and setting a benchmark within China's assisted driving sector.
Over the past year, XPENG's computing cluster delivered a 1,010% uplift in per-GPU training efficiency and a 4,360% gain in single-job training efficiency, while increasing hardware utilization from 40% to 90%, matching leading global AI standards.
The company unveiled its first blueprint for the world model, focusing on deliberative reasoning, controllable generation, and long-horizon forecasting to complement VLA2.0; together they form XPENG's physical-world AI system designed for real-world autonomous driving environments.
With VLA2.0's rollout underway, XPENG is accelerating deployment of AI applications including Robotaxi services and the IRON humanoid robot, slated for mass production by end-2026 and planned to serve as in-store guides from Q1 2027.
