13 September 2026
XPeng at IFA Berlin: Physical AI and the VLA 2.0 Turing Model
CleanTechnica's Larry Evans sat in on XPeng's IFA Berlin keynote with Dr. John MA: latency and safety as the constraints, Infini-VLA's 30-second memory, Streaming Inference, X-Foresight's 6-second prediction, the plan to distil VLA 2.0 onto single-Turing and dual-Orin cars, and robots riding in cars. With full credit and links to the original.

XPeng used IFA Berlin to talk about physical AI rather than another feature list, and CleanTechnica's Larry Evans was in the room: he sat through the keynote from Dr. John MA, XPeng's Senior Director and Head of ADAS & Smart Cabin Overseas Delivery, and then interviewed him afterwards. This is a summary of that piece, with full credit and links to the original below.
The framing: constraints come first
According to the presentation, the limits of AI in the physical world are what shape everything else. XPeng may have industry-leading compute in its Turing chipset, but onboard power is finite and has to stay energy-efficient; latency has to stay low; and low latency feeds into safety, which in turn feeds regulatory compliance.
Time is the centrepiece of 6.3.0
The biggest addition in VLA 2.0's 6.3.0 version is the time component, and the presentation broke it into four parts:
- Infini-VLA gives on-board decision-making a temporal context — what happened before, what is happening now — capped at 30 seconds of history to conserve compute;
- Streaming Inference then anticipates what comes next, which XPeng says improves response time by 300%;
- X-Foresight extends that to predicting the next 6 seconds, again bounded to keep compute in check;
- Flow Matching plots several candidate solutions and picks the best trajectory.
These are already running in XPeng vehicles in China, including its Robotaxi, which is operating in Guangzhou with wider coverage planned for next year. The IFA session was deliberately less technical than the Chinese 6.3.0 launch, since the Berlin audience was newer to the company.
The distillation question
XPeng currently ships VLA 2.0 on newer cars with three in-house Turing chips, and on its Robotaxi with a fourth for redundancy, with plans to reach single-Turing cars and then older dual-Orin models through a distilled model that needs less compute rather than losing features. Asked what owners will actually feel, John MA was careful:
"We still think that at least for safety. That must be guaranteed. There's no safety issue. And for the smoothness, maybe some slight difference..."
He added that public trials of the single-Turing machine have begun, Orin work is still in development, feedback is being gathered, and that any scenarios found lacking could be addressed by OTA.
Voice, languages and what onboard AI unlocks
Master Agent voice control — conversational control of vehicle functions, including driving — arrives on three-Turing Ultra cars and the Robotaxi. The interview touched on languages and dialects beyond English, Mandarin and Cantonese, including French, German, Spanish and Italian, with the hint that more news would come at the Paris motor show.
On running large models locally rather than in data centres, the answer was that a single VLA 2.0 architecture — vision, language and actions together — also serves as the foundation for XPeng's humanoid robots and flying cars, and that onboard-only decision-making has compliance and data-privacy advantages. The longer-term idea sketched in the interview is a closed loop between products: the car drives itself, the robot gets out to collect the shopping, and both return — described as a new "physical agent" rather than a digital one.
On the show floor
The booth featured the IRON humanoid robot, the flying car — which the report notes is being ramped towards production this year while seeking regulatory approval — and the L03. Visitors read the L03 as spacious and premium; several assumed power-closing passenger doors and were surprised by the price. Deliveries are expected to ramp in Europe, with VLA 2.0 activation there expected early next year, pending regulatory approval.
The piece closes with company scale — more than 31,600 employees, 44.79% of them in R&D — and a note that more is coming, particularly around humanoid robots.
Read and watch the original
- Original article: XPENG Brings Physical AI To IFA Berlin — Larry Evans, CleanTechnica, 7 September 2026 (includes photos by Larry Evans)
- Earlier deep dive on the 6.3.0 launch in China: XPeng Drives Physical AI To Next Level
- Presentation highlights video referenced in the article: YouTube
Source: CleanTechnica — "XPENG Brings Physical AI To IFA Berlin" by Larry Evans, 7 September 2026. All quoted passages and photographs belong to CleanTechnica and Larry Evans; this page is an independently written summary with attribution and links, not a reproduction of the article. Please read the original for the full interview.
Get the next one direct.
New articles, or a daily / weekly digest — the same list EVConduit keeps.
