XPeng VLA-Lite Distilled Version — First Impressions From Guangzhou

A few days ago in Guangzhou I got to experience XPeng's VLA-Lite distilled version. Some honest impressions to share.

A lot of people instinctively assume a "Lite" version is just the big model crudely chopped down — features cut, a shrunken edition. But after actually driving it, my feeling is different.

VLA-Lite's underlying technical paradigm is the same VLA 2.0 base as the full dual-Turing 6.3.0 version. It's not a separate, simplified architecture built from scratch. XPeng didn't take the crude "cut the model's capability" route — instead they maximally distilled the capability bandwidth of the 630.

Over a short ~40 minutes I didn't feel any essential difference from the 630 in the vehicle's cognitive ability or scene understanding. But the gap you can genuinely feel is concentrated in one place: decision latency.

The full 6.3.0's ultra-low latency was the single most memorable thing about this Guangzhou trip. The moment the car ahead moves, the vehicle reacts almost instantaneously — millisecond-level response — and the way it negotiates through traffic is crisp and decisive.

By contrast, this Lite distilled version running on a single Turing chip has a strong overall foundation, but the response-latency difference is clearly perceptible. The feel is somewhat like the 610 version on my P7: scene understanding and judgment logic stay consistent with the 630 — it's just that between "seeing" and "acting" there's a little extra time gap.

The 630's ultra-low latency is still a pretty obvious difference.

XPeng getting the VLA 2.0 distilled version running successfully means one unified tech stack can adapt across multiple tiers of compute hardware — from the triple-Turing and dual-Turing full-blooded flagships, all the way down to single-Turing vehicle platforms.

In the industry there are already companies running a separate architecture per model — multiple systems, multiple logic sets, multiple iteration and maintenance tracks that are all siloed from each other. Later upgrades and capability evolution are hard to sync. For these companies, the complex architecture eventually becomes "either don't update, or can't afford to update."

XPeng, meanwhile, runs a single base model and uses distillation to adapt downward across different compute tiers — so the entire product fleet can share the same "brain's" evolution. Every cloud model iteration lets multiple models gain cognitive upgrades simultaneously.

In the coming intense autonomous-driving competition, this is a crucial organizational capability: it lets the results of massive training-burn research efficiently empower the full product matrix, so more models can enjoy the benefits of model-based smart driving.

Of course, the dynamic-negotiation experience hit by latency remains the key optimization focus for the Lite version going forward. I look forward to future OTA iterations continuing to narrow this gap.

I haven't yet gotten my hands on the dual-Orin Max version's distilled experience.

---

Source: 电动兄弟 (Electric Brothers, Weibo) — translated & adapted for XPENGTech.