I think this left turn is very beautiful. This is actually a very large-distance left turn.
There is a small play. Yes, you can see that it just maneuvered a little between the two cars.
Yes, they drive to the front here. Drive to the front, and it will cut over immediately. Yes, very smart. It found that it cut immediately in advance. It won't hesitate, it won't stall.
Do you want to introduce yourself first?
Hello, my name is Ziqi. In fact, I have bought two generations of XPeng cars. From the earliest time, I was actually in the first batch to promote City NGP, and now I am the owner of this P7. In fact, I live in Foshan, so I deal with this kind of traffic commute every day. Basically, I will use this NGP during my daily commutes. According to the statistics in the app, basically 95% of the time is all covered by this NGP. So today I am also very honored and really looking forward to seeing this new version of NGP, the 5.3 Lite version of VLA Lite, and experiencing what's different from what I usually drive.
In fact, 95% is already a very high utilization rate.
Yes, but I guess VLA Lite can maybe improve it by two or three more points.
Then we got to the official route. Let's get on the road first.
I've tried a lot of systems, but I think XPeng has always been efficient. It didn't sacrifice efficiency just for some other subjective feelings. I think this is why I'm willing to use it. If it drives very slowly, then I guess I won't use it often. I think efficiency is more important, and second is peace of mind. You see, there was a pedestrian here just now, and it cut slowly from there, and we already had a pre-cutting process.
Yes, in fact, the whole feeling is very sensitive. The best autonomous driving, I think, is to drive very smoothly, drive very smartly, drive very safely in all kinds of complex scenes, and make you feel very stable across all kinds of environments—that is, whether you drive 10,000 times, it feels the same. Autonomous driving or L2 driver assistance has a great need to establish a sense of trust between people and cars. But with this sense of trust, I can have a thousand reasons to let you use it, but you only need one reason to not use it.
Yes, so letting it improve its ability to slow down smoothly and letting the whole model progress—this process of building a kind of natural rhythm is very important.
Yes, because in fact, I myself also love road trips. For example, when I go to Guizhou, in fact, I think in some sense it drives better than me because it knows the road. But when I went for the first time, I might hesitate at a certain intersection and wonder if I should go this way. I think this left turn is very beautiful; this is actually a very wide left turn.
We hope to retain some capabilities in the pruned and Lite versions. We don't want it to have a significant decline in safety and comfort in urban areas. We also hope to keep the feeling of the original VLA. This is what users like the most, and it's also what improves the safety baseline. But even with constraints, it means there are trade-offs in this process. The difference is that in the Ultra version, we hope it covers the entire scene, whereas in the Max version, it is aimed at the scenarios most users need, doing energy recovery and handling curves under limited conditions.
So I also want to ask Xianmin a question. I just said that we don't want to lose a lot of its abilities through model compression. Can you tell us specifically how you achieved that?
In fact, I'm also curious. First of all, we didn't do a crude reduction in the number of parameters by simply cutting the whole model in half or by a third. Instead, we made an architectural change at the base layer—that is, Hybrid ViT. This optimizes the computation mode across the entire model, allowing parts of the model to run with lower computing power and higher efficiency while preserving the overall backbone network. This is why the compressed version is now called Lite; it is not simply a straightforward model pruning. There are still many fundamental architectural innovations in the base layer.
For me, because my commuting distance is indeed relatively long, I would like it to feel more reassuring. I think a good driver assistance system should be like a cup of steady water. It won't constantly startle you or try to show off with dramatic maneuvers, but instead aligns with how everyone naturally drives. That matches my expectations.
Then this version should be quite in line with your expectations.
Don't you think this kind of lane change is much smoother than before?
Yes. I think the first time I rode in VLA last year, the biggest impression was, 'Oh my god, the way it changes lanes and decelerates feels even more comfortable than when I do it myself.' Because previously, after turning the steering wheel, there would be a sharp correction to counteract the movement. Now it's basically very smooth when it transitions.
The speed is also very fast. You can see it maintains about a 1.5-meter gap. This is what I just said—it provides both efficiency and peace of mind. In various scenarios, combining efficiency and reassurance makes you feel very stable.
This lane is actually right, because that side is the left turn lane. Yes, changing over in advance.
A smooth stop. It's quite comfortable to drive.
