XPeng's VLA 2.0 vs Horizon's HSD V2.1: A Reviewer's Split Verdict

A Guangzhou-based reviewer has published a head-to-head impression of the two most closely watched driver-assistance stacks in China right now — and reached a conclusion that will satisfy neither camp completely.

肉肉爸比ev, who also published the G9L launch press-conference record we covered, wrote on 21 September 2026 that a commenter under a video about Horizon Robotics' HSD V2.1 had asked him to compare it against XPeng's VLA 2.0. So he drove a Guangzhou route — "mainly Yuexiu, Liwan and Tianhe" — with both.

He is explicit about what this is: "three conclusions, representing only my personal view."

The three conclusions

"1. XOS 6.3.0's model capability is currently the closest to FSD — bar none. It is stronger than V2.1.

2. HSD V2.1 leads XOS 6.3.0 on functionality by roughly half a year. The model's direct-output reversing capability is showing its first results; parallel-parking escape usability is 100%. This time they also added a small restriction. By the end of the year, what will HSD be?

3. Don't say big compute is useless. That big compute is useful is an industry consensus. Some things are just for looking at — can 200 TOPS really dominate? You're joking."

He closes by saying that both systems' current experiences are "genuine 细糠" — slang for the refined stuff — and that some [rivals] "really shouldn't come near."

The split is the interesting part

Read those first two points together, because they cut in opposite directions and that is what makes the post worth recording.

On model quality, he gives it to XPeng — and in the strongest available terms, calling XOS 6.3.0 the closest thing to Tesla's FSD, "bar none," ahead of Horizon.

On functionality, he gives Horizon a six-month lead. Specifically he singles out reversing: HSD V2.1's model emitting trajectories directly for reverse manoeuvres, and what he reports as 100% usability for parallel-parking escape — squeezing out of a tight side-of-road bay. He also notes Horizon "added a small restriction" in this build, implying capability deliberately held back.

A reviewer who writes extensively about XPeng conceding a six-month functional gap to a competitor is a more credible signal than a clean sweep in either direction would be.

What HSD actually is

Worth stating plainly, because the abbreviation is doing a lot of work: HSD is Horizon SuperDrive, the end-to-end driving system from Horizon Robotics (地平线) — the Chinese chip company founded by Yu Kai (余凯), and not a Huawei product.

Horizon announced HSD V2.1 on 15 September 2026, at a Shanghai event marking its 15-millionth Journey (征程) chip coming off the line in a FAW-Volkswagen ID. AURA T6. The headline feature is what Horizon calls the first "all-scenario reversing" (全场景倒车): narrow-lane escape, extreme parking and dead-end U-turns, handled by end-to-end model trajectory output rather than stacked rules.

That builds on HSD V2.0 from June 2026, which Horizon described as "any point to any point, full-dimensional collision avoidance." Horizon also said HSD would reach mass production with a "leading new-energy vehicle manufacturer" within the year.

The relevant hardware runs from small to large: the Journey 6 series spans 10 to 560 TOPS, with Journey 6M at 128 TOPS bringing urban NOA down to the ¥100,000 segment — including to combustion-engined cars, since a single 6M can be air-cooled. The flagship Journey 6P is 560 TOPS. Horizon claims one in three Chinese cars with driver assistance uses one of its chips, more than 50% share of domestic-brand ADAS silicon, and second place in urban NOA share.

The compute argument, and the missing target

His third point is a culture-war shot rather than a technical one, but it lands in the middle of a real argument.

The industry has spent 2026 moving from peak TOPS to effective TOPS — how much of the nameplate figure a chip actually sustains while running a real model, at what power and what cost. XPeng leans on this framing hard, claiming a single Turing chip's effective compute is roughly equivalent to ten Orin-X chips. Horizon's counter-argument is the 128 TOPS Journey 6M doing one-stage end-to-end urban driving in a ¥100,000 car.

Against that backdrop, his line — "can 200 TOPS really dominate? You're joking" — is a defence of the big-compute camp, and of the 750 TOPS Turing approach specifically.

One caveat: he never names the 200 TOPS system he has in mind. It is not Journey 6M (128 TOPS) and not Journey 6P (560 TOPS), so it is either a rival configuration we could not identify or a comment-section claim rather than a shipping product. We have left the line as written and flagged it rather than guessing at a target.

Why "closest to FSD" is the sentence that matters

The FSD comparison is the frame Chinese reviewers reach for, and He Xiaopeng has been feeding it. At the G9L press conference he argued that XOS 6.3.0 "may not be a huge leap for users, but for the technology layer it is a huge leap," that the model's parameter count is larger, and that its narrow-road capability "may also exceed V14" — while noting pointedly that FSD is "very impressive in the United States, but not so impressive in other countries."

An independent reviewer calling XPeng's stack the closest domestic system to FSD — while still giving Horizon the functional edge — is the most useful shape that claim has taken so far.

What this review does and does not establish

What it is: one experienced owner-reviewer's subjective impressions, on one urban route in one Chinese city, on one day, explicitly labelled as personal opinion. There is no instrumentation, no repeat testing, no scoring rubric, and no blinding.

What it is not: a benchmark. Reverse-manoeuvre capability in particular is the easiest domain to over-read from a handful of successful attempts — "100% usability" in a post like this means "it worked every time I tried it," not "it works 100% of the time."

Route bias is real. Yuexiu, Liwan and Tianhe are dense central Guangzhou districts. That is a fair test of urban interaction, and no test at all of highway behaviour, rural roads, or other cities' signage and lane conventions.

Feature claims are Horizon's. The all-scenario reversing framing and the year-end mass-production claim come from Horizon's own announcement. This review corroborates that reversing works, not the scope of "all-scenario."

And the reviewer's orientation is worth knowing. 肉肉爸比ev writes about XPeng extensively. That does not make the conclusion wrong — it makes the concession to Horizon more notable, not less.

Source: 肉肉爸比ev, Weibo — translated & adapted for XPENGTech. Read the original post.

Related: XOS 6.3.0 begins rollout — VLA 2.0 Lite reaches Turing Max cars · Why dual-Orin XPengs can't run full VLA 2.0 · The G9L press conference, in full