XPeng 630 (XOS 6.3.0): More Than an OTA — the Underlying Upgrade, Explained

XPeng's 630 update is not a routine OTA iteration. This time the team spent much longer in the edit, because what 630 changes isn't a few scattered features — it changes the architecture underneath the second-generation VLA system.

The short version: XPeng's smart-driving stack is moving from understanding a 3D scene to understanding a 4D scene — space plus time. And that single change is what gives the system a genuine sense of memory.

From single frames to a "memory" of time

In the past, a neural network mostly understood the current frame. If a car ahead disappeared behind an obstacle, the tracking module could keep following its coordinates for a while — but once the target was fully occluded, the history broke. The system effectively forgot what had just happened.

The video's simplest example is a gap in the middle of the road. Single-frame detection sees "empty space." A car that is actually signalling and gradually turning to merge just looks like a slow-moving coordinate to the old system — so the car might be tempted to accelerate into the "clear" gap. The video calls this "spatial deception."

With the 630 update, the model no longer reads a single frame in isolation. VLA 2.0 extracts a temporal window of past events and represents them as a sequence, so the system sees a continuous action rather than a snapshot. It understands the before-and-after of a situation — and makes better passing decisions as a result.

What that actually changes on the road

The video walks through Guangzhou's hardest everyday scenes to show the difference:

• Loading/unloading and an excavator pulling out — instead of creeping forward into any gap, the system waits for the excavator to finish loading, reads when traffic behind the neighbouring car clears, and then moves out first.
• Continuous oncoming traffic on a narrow road — rather than diving into a gap and being "fooled to the right," the system slowly follows a forward opening, because it knows where each oncoming car came from and where it's going.
• Merging from a blind junction — launch timing, historically a weak point, now starts with a slow creep and completes the moment the bus alongside commits.
• Rush-hour cut-ins — the system can recognise an aggressive cut-in, anticipate a possible second cut-in, and instead of doing a defensive, efficiency-killing yield, it speeds up and completes an efficient pass.

Crucially, the video also checks whether 630 is simply "more conservative." It isn't. Compared back-to-back with the 610 version, 630 is more decisive on narrow roads and lane choices — not less. That's the hallmark of better scene understanding rather than safer-sounding timidity.

The catch: long memory is expensive

There's a catch to all this. A model's compute cost scales with sequence length. The more history you store, the more attention matching, feature extraction, and memory reads/writes are required. If every new frame had to be correlated against the full 30-second history, inference time would climb — "the more it remembers, the slower it reacts."

XPeng's answer is Streaming Inference. The video's analogy: old discrete inference is like a factory where every worker waits for all parts to arrive, assembles, then waits for the next batch. Streaming inference removes the idle waiting — while the first car is being assembled, the second is already getting its bolts, and the third is already being painted. Translated to the model: old results are cached and reused, so each frame only processes an increment, not the entire 30 seconds again.

The important detail is that memory and decision-making are combined. That's what resolves the natural conflict between "remember for a long time" and "react instantly."

A world model and 6-second prediction

On top of temporal memory, VLA 2.0 adds a world model and multi-probability prediction, with a prediction horizon of 6 seconds. In practice, every target gets more than one predicted outcome.

The experience changes in a subtle way. Before, the system guessed what another road user would do and made one plan to handle it. Now it pre-computes several contingencies — including the worst case — and keeps margin. In many scenes it intervenes earlier: a pedestrian suddenly appearing from a blind spot, a car door swinging open, a branch hanging across a blind corner.

The result is fewer heavy slowdowns and a more natural, less jerky feel, because high-risk scenes get optimised in advance rather than reacted to at the last moment.

What still needs work

The video is honest about the limits. Even after 630, the system's handling of two-wheelers remains insufficient. Compared with FSD V14, two-wheelers no longer behave as predictably as they did under rules-based or earlier models, and the difference is strongly felt.

There's also a meta-observation worth noting: from 610 to 620 to 630, XPeng has increasingly talked about architecture, technical choices, and generalisation rather than marketing individual scene-by-scene feature wins. That shift in language reflects what the foundation-model approach is actually buying.

One base model, many platforms

The final section covers cross-platform distillation. Rather than the industry's usual approach of crudely pruning the network or cutting core modules for cheaper hardware, XPeng trains the full VLA model in the cloud, then distils it down to a "Light" version for lower tiers.

Because the large-parameter base model carries strong scene-adaptation ability, only a small amount of local data is needed to adapt to new markets — the video points to Europe and Germany as examples.

So 630 isn't a single version update. It's a foundation-model-centric system that can push upward to Robotaxi, release downward to mass-production cars of different configurations, and generalise outward to overseas markets.

VLA 2.0 vs FSD V14

The video closes with the inevitable comparison. VLA 2.0 and FSD V14 are now in a direct, close-range matchup — but the underlying routes are different.

FSD V14's segmented model still prioritises North America's main hardware (HW4/HW5), and older hardware gets the model compressed to fit it, which fragments the experience across platforms. Its strong generalisation is built on a massive global fleet's data.

XPeng VLA, by contrast, was designed from the start to handle multiple hardware platforms, multiple markets, and cross-embodiment engineering — not just polishing the best result for the top-tier hardware. That, the video argues, is the core philosophical difference between the two foundation-model approaches.

---

Source: 飞机先生_ (Mr. Airplane), Weibo — video "不止是一次OTA！带你全面看懂小鹏630底层升级" (More than an OTA — a complete look at XPeng 630's underlying upgrade).