XPeng says its updated second-generation VLA driving model remembers up to 30 seconds of road context and raises end-to-end response speed by 300%. XPeng says the new XOS 6.3.0 software will make its global debut on the XPeng G9L, not on a BYD Han.
For a BYD Han owner, this is a useful view of how quickly Chinese manufacturers are moving their software stacks forward. It is not evidence that my 2020 China-market BYD Han will receive the same functions, because XPeng’s own announcement does not mention BYD or the Han.
What does XPeng’s VLA upgrade actually add?
XPeng says the first major update to its second-generation VLA model moves the system from understanding static 3D space to understanding dynamic 4D space and time. In plain English, XPeng says the model is meant to connect what happened before, what is happening now, and what may happen next rather than treating every moment as a separate image.
According to XPeng, the updated model uses a long-sequence architecture that can remember the previous 30 seconds of the road environment. XPeng says this gives driving decisions more context by connecting earlier actions, positions and road situations.
XPeng also says it has moved from discrete inference to continuous streaming inference. The company claims that this raises end-to-end response speed by 300%.
That is the useful bit behind all the AI language. A system that only sees one clean moment can recognise a car, pedestrian or traffic light. XPeng says its new model is designed to remember the situation developing around them.
Does XPeng XOS 6.3.0 apply to a BYD Han?
No, not from anything XPeng has published. XPeng says XOS 6.3.0 will make its global debut on the XPeng G9L, and its announcement does not say that the software, VLA model or its functions will be available for a BYD Han.
My BYD Han was bought new in China in 2020 and has been driven in Portugal since. It has CarPlay and Android Auto despite being a China-market car, as I wrote in my post about using CarPlay and Android Auto on a Chinese-market BYD Han. But cabin connectivity working on this car does not mean another manufacturer’s driving software can somehow arrive here too.
Different car. Different software platform. Different manufacturer.
And, obviously, different promises.
XPeng’s announcement is interesting because it shows where one Chinese manufacturer says its technology is heading. It does not create a roadmap for the Han.
How does XPeng’s updated model compare with its previous approach?
XPeng says the new VLA update changes how much context the model uses, how quickly it responds and how far ahead it predicts road events. These are XPeng’s own claims from its 1 September announcement, not independent test results.
| Area | Earlier approach described by XPeng | Updated approach described by XPeng |
|---|---|---|
| World understanding | Static 3D space | Dynamic 4D space and time |
| Road context | More focused on current information | Remembers up to 30 seconds of previous context |
| Inference | Discrete processing | Continuous streaming inference |
| Response speed | Baseline not stated | 300% faster end-to-end response, according to XPeng |
| Prediction | Not specified in the announcement | Predicts possible events up to 6 seconds ahead |
| Model size | Baseline not stated | End-side parameters increased 3.5 times, according to XPeng |
XPeng says its prediction model can use the position, speed, movement trend and road environment around other traffic participants to estimate what may happen in the next six seconds. It also says the updated VLA model uses a mixed architecture intended to reduce interference between tasks such as city driving, park roads and parking.
The company further claims its combined safety capability rises 20 times. That is XPeng’s characterisation of its own system, and the announcement does not provide an independent testing method in the text published on its site.
What does XPeng mean when it says AI understands time?
XPeng says “understanding time” means the model can remember the past 30 seconds, process the present road situation continuously and predict possible future movements. XPeng presents that combination as the difference between recognising objects and making decisions in a changing physical environment.
The examples XPeng gives are normal messy-road things: a car ahead braking suddenly, a pedestrian changing direction or another vehicle forcing its way into a lane. XPeng says the updated system can process these situations in parallel rather than following a simple see, calculate, output, then see again cycle.
This is where the announcement gets more interesting than a normal software-version number. A bigger screen animation or another voice command is easy to understand. A model remembering half a minute of road context is harder to see from the driver’s seat, but it is the part XPeng says changes the underlying driving logic.
Whether it works as claimed is a separate question. XPeng published the claim. Real roads usually publish the review later.
What new car controls does XPeng say it is adding?
XPeng says the update introduces a vehicle-wide controller called Master Agent, combining its VLA driving model with a vision-language model for driving and cabin functions. XPeng says the system is intended to understand a user’s broader intention, split it into tasks and coordinate separate agents for driving, chassis, cabin and body control.
XPeng also says the new version adds voice-controlled roadside parking and voice-controlled nearby parking. According to XPeng, a driver can give a voice parking instruction while the vehicle is in its driving-assistance state, and the vehicle can look for a suitable place and complete the parking manoeuvre.
That is a proper step beyond saying “set temperature to 22 degrees”. XPeng is describing a system that takes an instruction and then carries out several actions.
Still, it is XPeng describing its own target experience. The announcement is not a test drive, and it is not a promise for every XPeng already on the road.
When will XPeng bring the second-generation VLA to Europe?
XPeng says it recently completed local acceptance testing for its second-generation VLA in Germany. The company says the model, trained on Chinese data, delivered a real-world experience on German urban roads that was highly similar to its domestic experience without a significant increase in local training data.
XPeng says its target is to obtain regulatory approval in Europe in the first half of next year and then gradually deliver the second-generation VLA to overseas users. That is XPeng’s target, not a confirmed European launch date.
For owners in Portugal, that wording matters. “Target” is not “available”, and “overseas users” is not a list of countries, models or existing cars. XPeng has given a direction here, not the final small print.
FAQ
What is XPeng’s second-generation VLA update?
XPeng says it is the first major update to its second-generation VLA model, designed to move from static 3D spatial understanding to dynamic 4D understanding of space and time.
How much road context does XPeng say the new model remembers?
XPeng says its new long-sequence architecture remembers the previous 30 seconds of the road environment and uses that context in driving decisions.
Does XPeng say the update responds faster?
Yes. XPeng says its streaming inference approach raises end-to-end response speed by 300% compared with its previous approach.
Will XPeng XOS 6.3.0 come to a BYD Han?
XPeng does not say that. XPeng says XOS 6.3.0 will make its global debut on the XPeng G9L, and its announcement does not mention BYD Han compatibility.
What parking functions does XPeng say the update adds?
XPeng says the update adds voice-controlled roadside parking and voice-controlled nearby parking, allowing a vehicle in its driving-assistance state to find a suitable parking location and complete the manoeuvre.
When does XPeng expect to bring the second-generation VLA to Europe?
XPeng says it aims to obtain European regulatory approval in the first half of next year and gradually deliver the system to overseas users. The company does not give a confirmed launch date, country list or model list in this announcement.
Source
This post comments on a single primary source: the company news announcement published by XPeng on 1 September 2026. XPeng published the original page in Chinese; the English rendering of its title, technical descriptions, dates and claims in this post is my own translation. Every statement here about the VLA update, XOS 6.3.0, its claimed figures, the G9L and XPeng’s European plans is XPeng’s account, not mine.
Leave a comment