XPeng ships VLA 630: on-device driving model up 3.5×, 3× faster
Chinese electric-vehicle maker XPeng is pushing the self-driving model arms race from the datacenter down into the car, announcing a major upgrade to its second-generation visual-language-action (VLA) model that runs entirely on-vehicle.
XPeng unveiled the first big update to its second-generation VLA, version 6.3.0 — dubbed the "630" — with the on-device model tripling its parameters and its end-to-end response getting three times faster. According to the company, the in-car model now packs 3.5 times the parameters of the previous release and roughly 15 times that of a typical industry VLA system, while end-to-end response speed jumps about 300 percent. The upgraded system makes its global debut on the XPeng G9L, with full details expected at the company's Physical AI sharing event on August 27.
The headline number is the on-device parameter count, but the real story is whether a much larger model can still think fast enough inside a car. Cloud models can wait an extra few hundred milliseconds to consult a GPU farm; a vehicle at speed covers several meters in that same window, so a bigger model that infers slowly is worthless or worse. XPeng says it rebuilt the VLA around its proprietary Turing chip, co-optimizing model, silicon, and compiler so the larger network still runs in real time — the difference between a model that merely recognizes something ahead and one that smoothly sequences slowing, yielding, and lane changes.
This is the front line of the assisted-driving competition against Tesla, Huawei, Li Auto, and NIO, who are all chasing the same end-to-end path. The VLA bet folds visual understanding, behavior prediction, and control into one model instead of a brittle stack of rule-based modules. What will decide the race is not the parameter figure XPeng quoted but whether the 630 drives reliably in everyday stop-and-go traffic and, crucially, whether the update reaches existing Max vehicles through distillation rather than staying locked to a handful of new models.
What to watch: whether the August 27 event surfaces real-world road data and a rollout schedule for existing vehicles — that will tell us if the 630 is a genuine productization step or a launch-vehicle showcase.
Autonomous-driving specs are easy to inflate and hard to verify on a test drive — what metric would convince you a driving model has actually gotten better? Tell us in the comments.
