XPENG’s VLA 2.0 vs Tesla FSD: Claims, Controversy, and What the Tests Actually Show

XPENG’s latest autonomous‑driving software, VLA 2.0 (Vision‑Language‑Action 2.0), has sparked debate after the company promoted a comparison showing it outperforming Tesla’s Full Self‑Driving (FSD) on Chinese city roads. The claim centers on a simple metric: how often the human driver had to intervene. But the test—and the conclusions drawn from it—remain controversial.

XPENG’s headline claim: fewer driver takeovers

In a promoted comparison conducted on urban roads in China, XPENG said a P7 equipped with VLA 2.0 required only two driver takeovers, while a Tesla Model 3 using FSD required seven on the same routes.

The company presented this as evidence that its system performs better in dense Chinese traffic conditions. Driver takeovers are a common metric in advanced driver‑assistance testing because they indicate moments when the automated system cannot safely continue without human input.

However, the comparison provided limited public detail about factors such as:

• exact route length and complexity
• traffic conditions
• software versions used
• whether the runs were repeated
• independent validation of results

Because of those missing details, the claim is best viewed as a demonstration rather than a controlled benchmark.

Why critics say the comparison may not be fair

A major criticism is that Tesla’s FSD in China currently runs a more restricted version than the latest releases used in the United States.

Reports indicate that the comparison involved FSD Version 13 in China, while Tesla has continued advancing the system elsewhere.

That difference matters because localization and regulatory constraints can limit capabilities such as:

• mapping data availability
• training data specific to Chinese traffic behavior
• feature rollout timelines

As a result, some analysts argue that comparing a China‑native system optimized for local roads against a restricted regional version of Tesla’s software may not reflect the true global capability gap.

The core idea behind VLA 2.0

VLA 2.0 represents XPENG’s attempt to rethink the architecture of autonomous driving AI.

Traditional driver‑assistance systems often follow a staged pipeline:

1. perception (detect objects and lanes)
2. scene understanding
3. planning and decision‑making
4. vehicle control

XPENG says VLA 2.0 collapses much of this structure into a single end‑to‑end model that directly converts visual input into driving actions.

The system removes the traditional "language translation" step used in many multimodal AI pipelines and instead generates actions directly from perception signals.

XPENG describes the model as a “physical‑world large model” designed to understand and act in real driving environments.

Scale: data and compute behind the system

According to company materials and industry reports, the VLA 2.0 training system includes:

• about 72 billion parameters in its cloud model
• training on roughly 100 million driving video clips
• on‑vehicle compute reaching roughly 2,250 TOPS

XPENG says this dataset represents the equivalent of tens of thousands of years of human driving experience, though such figures come from company estimates rather than independent verification.

The system was rolled out via over‑the‑air updates starting in March 2026 for several XPENG models, including the P7, G7, and X9 in higher trims.

Real‑world impressions from reviewers

Independent test drives suggest VLA 2.0 is a serious contender in the advanced driver‑assistance race.

During one demonstration drive in Beijing, a reviewer navigated heavy city traffic for around 40 minutes without needing to intervene once, highlighting the system’s smoothness and confidence in complex environments.

Other reviewers also described VLA 2.0 as impressive but noted limitations compared with Tesla’s system. Some observed that:

• Tesla’s FSD still appears more mature overall
• XPENG’s system can rely heavily on lane‑level mapping
• certain edge cases remain less proven

Overall, reviewers tend to frame VLA 2.0 as a rapidly improving rival rather than a clear leader.

XPENG’s public timeline for beating Tesla in China

XPENG’s leadership has been unusually explicit about its ambitions.

Founder and CEO He Xiaopeng has publicly stated that XPENG aims to surpass Tesla’s self‑driving capability in China by August 30, 2026.

The goal is to match the driving experience delivered by Tesla’s newer FSD versions in the United States—specifically the level of performance seen in systems such as FSD V14.2 tested in Silicon Valley.

Beyond that milestone, the company has suggested that VLA‑based systems could eventually support Level 4 autonomous driving, though timelines remain uncertain and depend heavily on regulatory approval and further technical validation.

The bigger picture: a fast‑closing gap

Tesla has long been seen as the benchmark in consumer self‑driving technology. But the rapid progress of Chinese EV makers like XPENG is changing that narrative.

The VLA 2.0 comparison may not be a definitive test, yet it highlights a broader shift: autonomous‑driving development is no longer dominated by Silicon Valley alone.

Chinese manufacturers are leveraging massive driving datasets, rapid software iteration, and aggressive AI development cycles—factors that could reshape the global race toward higher‑level autonomy over the next few years.