XPeng held what it called a Physical AI and second-generation VLA experience day on Thursday, and the headline number is the kind that’s easy to misread at first glance. The company says its updated driving model now holds onto 30 seconds of what a car has just driven through and uses that to work out what’s likely to happen over the next six seconds, before it happens.
That’s the core idea behind X-Foresight, a new predictive layer XPeng is adding to its second-generation VLA (Vision-Language-Action) system. It’s part of a broader update, version 6.3.0, that XPeng is now rolling out, timed to debut alongside the global launch of the G9L SUV.
What Actually New
XPeng splits version 6.3.0 into three parts, organised by tense rather than by feature list. The past is handled by Infini-VLA, which builds a rolling 30-second memory of what the car has just seen and folds it into current decisions. The present runs on what XPeng calls streaming autoregressive reasoning, essentially continuous inference rather than a fixed decision cycle. The future is where X-Foresight comes in, generating several possible outcomes for the next six seconds and simulating how nearby drivers and pedestrians might behave, so the car can pick a response before a situation fully develops rather than reacting once it already has.
Which XPeng cars support the new update?
XPeng has only committed the full 6.3.0 update to vehicles with its triple-Turing-chip hardware, such as the G9L, G7 Ultra and new P7.
On paper, the raw numbers back up the framing. The on-device model’s parameter count rises 3.5 times over the previous version, matching what XPeng told investors on its earnings call three days earlier. End-to-end response times improve by roughly 300%, and XPeng also introduced an MoT hybrid architecture that hands different tasks (city driving, campus roads, parking) to separate specialised parts of the model instead of one system trying to cover everything at once, which the company says makes switching between scenarios more stable.
| What’s new | Detail |
|---|---|
| Prediction window | Simulates up to 6 seconds ahead (X-Foresight) |
| Memory window | Retains 30 seconds of recent driving (Infini-VLA) |
| Parameter count | Up 3.5x over the prior on-device model |
| Response speed | Up roughly 300% |
| Claimed safety improvement | 20x, a composite of six separate measures |
| Rollout starts | Late August 2026, alongside the G9L launch |
| Hardware required | Triple-Turing-chip vehicles only, no confirmed backward compatibility |
About That Twentyfold Number
The twentyfold safety claim is the figure most likely to end up in a headline, and it deserves a closer look before it does. XPeng describes it as a multi-dimensional comprehensive safety capability improvement, built by combining six separate measures: risk identification, trajectory planning, handling of extreme conditions, response latency, future projection, and a category XPeng lumps together as generalisation and robustness. The company hasn’t published the individual scores behind those six categories or explained how they’re weighted into a single multiplier. That doesn’t make the number meaningless, but it does make it closer to a marketing summary than something you could independently verify or compare against a rival’s safety claim.
A Chip That’s Finally Doing Something
A second announcement from the day was arguably the more concrete one: XPeng confirmed that the third of its in-house Turing chips, hardware that’s been sitting in customer vehicles since November’s reservation opening for the G7 Ultra and the new P7, is now active and running a system called Master Agent. It handles instructions by voice: pulling over, finding nearby parking, loosely worded navigation requests, and multi-step planning that used to require several separate commands. XPeng summed it up as “one brain runs the car, one voice controls it.”
That’s a meaningful shift from the chip sitting unused. Owners who paid for the third-chip hardware upgrade, installed under the front passenger floor without any need to strip the car down, have effectively been carrying 750 TOPS of spare compute waiting for software to catch up to it. Thursday is when that software showed up, at least for cars built with all three chips.
So When Does This Reach Europe?
XPeng hasn’t given a date for 6.3.0 specifically outside China, and there’s a reasonable explanation for that gap once you separate two different kinds of update the company issues. General comfort and infotainment updates move fast: I covered XPeng’s XOS 5.9.5 rollout last week, an update that reached 64 markets within a few months of finishing beta testing, cooling features, cabin filtration, voice command changes, nothing that needs a regulator’s sign-off before it ships. Assisted-driving software runs on an entirely different clock, because every market has its own rules about who’s liable when the car is making the decisions.

XPeng told investors it’s aiming to secure regulatory approval for the base second-generation VLA architecture in Europe in the first half of 2027, with a wider overseas rollout to follow from there. That target already assumes months of type-approval work still ahead: He Xiaopeng personally completed local acceptance testing for the system in Germany back in July, and that was for the underlying architecture 6.3.0 builds on, not the new version announced this week. If the base system is still working through European approval on a 2027 timeline, the specific improvements layered on top of it in 6.3.0 have no realistic path to arrive here any sooner.
There’s also a pattern in XPeng’s own recent history that applies here. When I looked at the new-generation P7 that’s currently China-only, the pattern was consistent: XPeng splits its most capable smart-driving hardware into paid tiers at home, and features that make the trip to Europe tend to arrive simplified rather than in full. Master Agent’s voice-command layer, in particular, looks like exactly the sort of feature that could see a slower or trimmed European rollout, since natural-language driving commands need language-specific tuning that a straightforward OTA push doesn’t solve on its own.
What is XPeng’s Master Agent?
Master Agent is a voice-controlled system running on XPeng’s third in-house Turing chip that lets drivers issue instructions like pulling over, finding parking or complex multi-step requests through natural speech
Put those things together and a reasonable estimate looks like this: the underlying architecture behind 6.3.0 might clear European regulatory approval around the first half of 2027, at the earliest, and whatever version of it actually ships here is more likely to be a trimmed-down variant than the full feature set shown off in Guangzhou this week. None of that is confirmed by XPeng, but it lines up with how every comparable rollout from the company has played out so far. For context on how far these systems are legally allowed to go regardless of what the marketing implies, my explainer on SAE’s autonomous driving levels is a useful starting point.
The Bottom Line
Six-second prediction and a 30-second memory window mark new territory for driver assistance software, and the underlying architecture shift from a reactive to a forecasting system is a real technical change, not just a bigger number on a slide. The twentyfold safety claim is a different story, an unverifiable composite figure that says more about how XPeng wants this update covered than about what it actually delivers. And for European readers specifically, the most useful takeaway is patience: this is a China-first rollout tied to a single new SUV launch, running on hardware that isn’t standard even in XPeng’s own European lineup yet, with the underlying software still working through the regulatory process that determines when, and in what form, it eventually gets here.
Marko Lubar is the Founder and Editor of ElectricFleet Online, where he covers the latest developments in electric vehicles, battery innovation, autonomous driving, artificial intelligence, and the technologies shaping the future of mobility.
Learn more about Marko and the mission behind ElectricFleet Online on the About Us page.









