XPeng's Distilled VLA Reaches Dual-Orin MAX Cars: First Impressions

For XPeng owners on the older dual-Orin (双O) MAX hardware, the VLA era has been an uncomfortable wait. The company's newest assisted driving is built for its self-developed Turing chip, and a single Turing is widely assumed to out-compute a pair of Nvidia Orin-X boards for this kind of workload. The fear in those owners' comments was not really about silicon, though — it was about attention.

The original post behind this article collects some of them. Translated:

"Whatever the upgraded assisted driving ends up like, at least there's the attitude — they've touched us dual-O owners."

"Dual-Orin is a general-purpose chip; the compute you can actually use is half of it, and inter-chip bandwidth is orders of magnitude below on-chip communication. Don't expect XPeng to spend much on an old architecture where the money's already banked. Anyway, every maker's dual-Orin assisted driving is rubbish right now."

"I think it's good. Slow doesn't matter — it can be optimised later. Latency is the key: I hope dual-O latency can reach this level too. Less latency means far fewer unnecessary hard brakes."

The first taste

Auto photographer 铁皮摄影师 (Tiepi) went to a first taste of the dual-Orin distilled build together with 肉肉爸比ev, and came away arguing that this matters more than the single-Turing MAX build. In his words:

"The dual-O distilled version finally gets its first taste. Today I went tasting again with 肉肉爸比ev, and to me this dual-O build is more significant than single-Turing MAX. After my single-Turing MAX video, a lot of dual-O owners in the comments clearly didn't dare expect much from their own hardware platform — they felt dual-O compute can't match single Turing, and worried that XPeng's engineers wouldn't put much effort into polishing it. But today's taste should give everyone some peace of mind: there is no question of compute limits, or of going through the motions. This model is strong — strong enough that 肉爸 commentated excitedly the whole way and I couldn't stop clapping."

The claim that matters: alignment with Ultra's 630

The headline finding is a comparison. Across the vast majority of scenarios — including some that were fairly extreme, or simply odd — Tiepi says the dual-Orin build's model capability can align with the Ultra's 630 version, the build running on XPeng's Turing-equipped flagship hardware.

He is careful about what that does and does not mean. It is still an engineering build that needs polishing. But the result, he argues, is enough to retire two worries that had been circulating for months:

• that dual-Orin cars would be compute-limited in a way owners would feel every day; and
• that the "distilled" version would be a token trim — a model cut down just to say the platform had been served.

Instead, he describes it as a system carefully tuned for the dual-Orin-X platform specifically, rather than a port of something built for other hardware.

Why the coverage argument matters as much as the build

The second half of the post widens the lens. XPeng's dual-Orin MAX cars are not a small fleet — the hashtags on the original post point to MONA M03 / MONA L03 owners being able to upgrade to the second-generation VLA, and to He Xiaopeng joining Max owners for their first experience of the VLA distilled version.

If that installed base can genuinely run the distilled model at a level close to the flagship build, the argument goes, the reach of high-end AI assisted driving moves up another tier — not because a new flagship shipped, but because a large existing fleet was lifted. Tiepi's phrasing: when the dual-O MAX cars can do this, "the coverage of genuinely high-end AI assisted driving will step up to a new level".

The post was reposted with the summary line that gives the mood of the day: "Teacher Tiepi shouted out — this is smart driving, this is physical AI. With dual-O like this, how could you not love it?"

Two caveats worth keeping in view. This is a first-impressions account of a non-final engineering build, not a measured comparison; and it comes from invited testers rather than an independent test. The useful takeaway is narrower but real: the distillation path to older XPeng hardware is being worked on seriously, and the first reports from the sharp end are positive.

Source: 铁皮摄影师, Weibo (reposted by xp-小宇宙) — translated & adapted for XPENGTech.