What is VLA 2.0?
XPeng rolled out VLA 2.0 via over-the-air updates in March 2026, delivering it to the P7, G7, and X9 in their "Ultra" configurations. The architecture represents a fundamental break from the company's earlier NGP system.
Instead of chaining separate perception, planning, and control modules, VLA 2.0 uses a single end-to-end vision-to-action neural network. The cameras see; the model decides; the car acts. By eliminating the intermediate translation layers, XPeng claims it has cut latency and reduced hard braking events by 99 percent while improving driving efficiency by 23 percent over the previous generation.
The compute comes from XPeng's proprietary Turing AI chip, which the company says delivers up to 2,250 TOPS of sparse compute in production vehicles. The model was trained on 100 million clips from what XPeng calls "extreme driving scenarios" — and according to independent tests, that claim is proving easier to believe.
Beijing: the world's toughest driving exam
Beijing traffic is not a polite queue. Lane markings are suggestions. Merging is Darwinian. If an autonomous system hesitates, it will sit at an intersection until the battery dies.
During a 40-minute test route, VLA 2.0 navigated complex intersections, managed aggressive cut-ins, and kept pace with traffic without being either overcautious or dangerously assertive. The standout moment came during a heavy-traffic lane change: instead of freezing or handing control back to the human, the system committed to a tight gap the way an experienced local driver would — firmly, smoothly, and successfully. It was nerve-wracking to watch, but it worked.
"You can tell this system was trained for Chinese road conditions," the test driver noted. "It doesn't drive like a cautious American suburbs algorithm dropped into Beijing chaos. It drives like it belongs there."
The inevitable Tesla comparison
XPeng CEO He Xiaopeng has been openly measuring his company against Tesla. Last year he flew to Silicon Valley and spent five hours testing FSD v14.2 in San Francisco, later calling it "near-Level 4" performance — high praise from a rival. Then he set a deadline: XPeng's VLA system must match FSD v14.2's performance in China by August 30, 2026. He even made a tongue-in-cheek bet with his head of autonomous driving: miss the deadline, and the executive has to run naked across the Golden Gate Bridge.
Based on the independent Beijing test, the August target looks not just achievable, but perhaps already met. The system reportedly felt comparable to Tesla's latest FSD v14 — a comparison not made lightly by those who have driven both.
Yet there is a crucial asymmetry. Tesla's newest FSD software is still not approved in China. Chinese Tesla owners with advanced driver-assist features are stuck on v13, while North American customers already run v14. Beijing has repeatedly delayed full FSD approval, and Musk's February 2026 prediction of imminent clearance never materialised. That regulatory gridlock gives XPeng a rare window to establish itself as the on-the-ground leader in the world's largest EV market.
Why Europe should pay attention
The story is not just about China. XPeng has already planted flags in Europe: the brand entered Norway in 2021, launched in Germany in 2024, and began selling the G6 and G9 in Italy in 2025. More importantly, cars for the European market are now being manufactured in Graz, Austria, in partnership with Magna Steyr — a move that sidesteps tariffs and speeds up homologation.
But the real European vector may be Volkswagen. The German giant owns a 4.99 percent stake in XPeng and has already paid over €200 million in technology licensing fees. VW is the first external customer for VLA 2.0, deploying it in a new Chinese-market electric SUV. The two firms have also signed an expanded agreement to integrate XPeng's electronic architecture across VW's platforms — including ICE and PHEV models in China. It is not speculation to say that this collaboration could eventually put XPeng-derived autonomy into VW Group vehicles sold in Frankfurt, Milan, or Amsterdam.
That matters because Europe's regulatory landscape for autonomous driving remains far more restrictive than China's. The EU's UNECE R157 regulation currently permits only hands-off Level 3 systems — such as Mercedes-Benz Drive Pilot — under strict speed and road-type limits. Tesla's FSD Supervised, which operates as an advanced Level 2 system in the United States, is not available in the European Union in anything close to its North American form. If VW can bring a UNECE-homologated VLA derivative to market, it could leapfrog both Tesla and Mercedes in the race for scalable European autonomy.
The field is getting crowded
XPeng is not the only Chinese challenger. BYD now ships its "God's Eye" system as standard on vehicles priced around €28,000. Huawei has pledged $11–13 billion for autonomous driving software over the next five years. Xiaomi is pushing smart driving hard in its SU7 lineup. All of these systems are iterating rapidly in China, the world's largest EV market, while Tesla's latest software remains grounded there by bureaucracy.
The competitive moat Tesla enjoyed for years is getting shallower by the month. What once looked like an insurmountable lead in real-world neural-network training is now being challenged by rivals with equally massive datasets, dedicated AI silicon, and — in XPeng's case — a willingness to ship end-to-end models that actually work in the world's most chaotic traffic.
The subscription question
There is also a business-model divergence worth watching. Tesla has moved to a subscription-only pricing model for FSD in the United States, charging $99 per month and signalling that the price will rise as capabilities improve. The one-time purchase option was eliminated earlier this year.
XPeng, for now, includes VLA 2.0 in the purchase price of its Ultra vehicles. The company has hinted it may charge for the service eventually, but the current message is clear: advanced autonomy is a feature of the car, not a recurring tax on the owner. In a European market already sensitive to total cost of ownership, that distinction could prove decisive.
What does VLA stand for, and why is it different from XPeng's older NGP system?
VLA stands for Vision-Language-Action. Unlike the previous NGP system, which used separate modules for perception, planning, and vehicle control, VLA 2.0 is a single end-to-end neural network that translates camera input directly into driving decisions. This reduces latency and eliminates the errors that can occur when information is passed between discrete software layers.
Will XPeng VLA 2.0 be available in Europe?
Not immediately. XPeng sells cars in Norway, Germany, and Italy, but VLA 2.0 is currently deployed only in China. However, Volkswagen — which holds a stake in XPeng and licenses its technology — could eventually bring a UNECE-homologated derivative to European VW Group models. XPeng is also now manufacturing vehicles for Europe in Graz, Austria, suggesting a longer-term commitment to the market.
How does this compare to Mercedes-Benz Drive Pilot, the only Level 3 system approved in the EU?
Mercedes Drive Pilot is a true Level 3 system certified under UNECE R157, meaning the driver can legally take eyes off the road in specific conditions — but only below 60 km/h on suitable motorways. XPeng VLA 2.0, like Tesla FSD, operates as a Level 2 system where the driver must remain engaged. Its strength is broader capability across complex urban environments rather than legal liability transfer at low speeds.
Source: https://electrek.co/2026/04/29/xpeng-vla-2-test-drive-tesla-not-alone-full-self-driving/