XPENG's 2nd-Gen VLA Upgrade: Introducing 30-Second Temporal Memory and 6-Second Prediction
nashnova research
XPeng released the first major upgrade to its second-gen VLA, enabling the driving model to recall 30 seconds of road history and predict 6 seconds ahead; the G9L will be the first vehicle to carry it. Separately, XPeng's Robotaxi has secured driverless testing approval in Guangzhou — production cars and autonomous validation are now on converging tracks.
What actually changed in this upgrade?
The core addition is a time dimension. Infini-VLA — a long-sequence architecture that lets the model remember recent road conditions — draws on the past 30 seconds of driving context. X-Foresight, a prediction module, projects scene changes 6 seconds into the future.
This means → the model no longer decides frame by frame. It carries memory and anticipation — critical for scenarios like a pedestrian doubling back, a lead car slamming its brakes, or a side vehicle cutting in.
XPeng says on-device model parameters grew 3.5×, response speed improved 300%, and overall safety capability rose 20×. Streaming inference tightens the loop between perception, computation, and vehicle action.
Where does the Robotaxi stand now?
XPeng's Robotaxi, running second-gen VLA, recently obtained remote-testing certification in Guangzhou. It can now conduct driverless road tests — no safety operator in the driver's seat — on tier-one, -two, and -three test roads.
This means → XPeng has entered the critical phase of truly unoccupied-driver-seat validation, moving past the earlier stage where a safety operator sat beside the wheel.
XPeng plans to feed some Robotaxi capabilities back into production cars — voice-activated pull-over, nearby parking, and more. In plain terms = features proven on an unmanned taxi gradually migrate to the car you can actually buy.
Is the G9L the delivery vehicle for this upgrade?
The G9L opened pre-sales on August 11 at RMB 259,800, available in both pure-electric and range-extended versions.
XPeng also plans to link the driving model with the cabin model on the G9L, letting voice commands trigger driving actions directly.
The production version remains classified as combined driving assistance — the driver must stay attentive. "L4 capability trickle-down" describes where the technology came from, not the regulatory level of the feature. Real-world takeover rates and user feedback will be the true test of whether this architecture upgrade translates into a noticeable experience gap.
What are competitors doing?
Li Auto has already pushed VLA driving features via OTA. Huawei launched ADS 5 in April, incorporating cloud-based multi-agent training, online reinforcement learning, and on-vehicle risk scenarios.
The branding differs, but the direction is the same: away from rule-based and modular stacks, toward systems driven by data + compute + large models together.
This reflects an industry that has entered a model-capability race. A single launch-event demo is no longer a moat — sustained data collection, simulation validation, and iteration speed determine the ceiling.
Can XPeng's finances support the pace?
Q2 2026 R&D spending hit RMB 2.91 billion, up 32.1% year-on-year. Revenue in the same quarter reached RMB 19.74 billion, up 8%, with a blended gross margin of 20.7%.
As of end-July, XPeng's cumulative deliveries topped 1.2 million vehicles. In plain terms = more cars on the road means more real-world driving data feeding back into model training.
R&D investment is accelerating, but revenue and margins are improving in step. This means → XPeng currently has the financial room to sustain heavy autonomous-driving R&D — the question is how long that balance holds.
市场有风险，内容仅供研究参考，不构成投资建议。