XPeng's Dual-Orin VLA 2.0 Distilled Build: First Drive in a MONA M03

XPeng's dual-Orin (双O) distilled VLA 2.0 build has had its first proper outing, and the review that matters most for owners on that hardware is a two-hour drive in a MONA M03 by 肉肉爸比 — comparing it throughout with the three-Turing "630" build.

▶ Play the 11-minute preview (English subtitles burned in)

The same video is on our YouTube channel: XPeng's Dual-Orin VLA Distilled Build — First Drive on MONA M03 (English subtitles).

What it does in the preview

The build is early — the reviewer says up front to ignore the SR display state and judge the driving. What follows is a fairly complete tour of the behaviours dual-Orin owners have been waiting to see:

• Leaving an underground car park by itself, from a standstill, into roaming mode, out through the exit barrier — with the reviewer noting at one point how close it passed another car ("millimetre-level").
• Continuous oncoming traffic on a narrow road, where a white car squeezes past a tricycle: the model doesn't brake hard, just eases off slightly and passes calmly.
• A pedestrian pop-out, which it handles — though he wants the deceleration amplitude increased, because a pedestrian's crossing path can't always be predicted.
• Long temporal sequence in action: waiting, then committing, at a junction with oncoming traffic, a big truck and a distant car. Watching it decide whether to go after the truck clears is the sequence he uses to make the point.
• Efficiency-first lane choices: reading a slower bus, two quick flicks of the wheel and a lane change to overtake; skipping a U-turn queue by moving right.
• Traffic-light pre-start: moving off with three and then four seconds still on the countdown — frequent enough that he calls it emergence rather than a one-off.

His summary line is the one that will travel: "Am I actually driving the Orin distilled build right now, or the three-Turing version?"

Where it differs from three-Turing

He is careful not to oversell it. The gap shows up mainly in reaction: the dual Orin-X build observes for longer before committing to a lane change, while three-Turing anticipates sooner and looks more composed — "effortless" on these scenes, where dual Orin-X is noticeably less decisive.

In pure model capability, he says, daily driving often won't feel very different. The build's ceiling, smoothness, comfort and flexibility are its clear strengths. Over a two-hour drive the only real problem exposed was one lane-selection error.

His verdict: if the production version retains 80% of this, he'd be satisfied.

How XPeng says the distillation works

The video closes on XPeng's own description, given by Liu Xianming at the XOS 6.3.0 launch:

• It is not simple parameter cutting or pruning to fit weaker chips;
• XPeng redesigned capability transfer across the student architecture, teacher model, base capability and training data;
• a new encoding scheme keeps only key information, heavily compressing the token count so the compute load drops;
• two-layer supervision ensures important information isn't cut;
• for dual Orin, the work is harder — cross-platform, cross-architecture transfer — and has required optimisation down to architecture and operator level.

That description matches what the drive shows: not a reduced version of VLA 2.0, but the same behaviour with a slightly slower decision clock.

Timing

The Chinese post carrying this video notes that 烟隐's reading of the available information is that the dual-Orin rollout could arrive in November. That is speculation, not an announcement, and it concerns the China rollout — nothing official has been said about timing for other markets.

Source: 肉肉爸比, Weibo — reposted by 烟隐烟隐, 15 September 2026. Translated & adapted for XPENGTech.