4 September 2026
XPeng's 2nd-Gen VLA Upgrade: From Understanding the World to Understanding People
A plain-language summary of VLA 2.0's first major upgrade: 30s memory, 6s prediction, 300% faster response, and the Master Agent.
Today XPeng held its Physical AI sharing and 2nd-gen VLA new-version experience day. Big news: the third Turing AI chip is lit, and 2nd-gen VLA gets its first major upgrade since mass delivery.
In short, the upgrade moves AI from "understanding the world" to "understanding people."
Cockpit: Xiao P swaps its base model architecture and gains hearing, vision, reasoning, and expression. It's no longer "you say one thing, it executes one thing" — it understands intent, invokes vehicle capability, and completes tasks. A vague instruction becomes navigation; a complex trip gets planned proactively. With VLA + VLM cockpit integration, the on-car agent does cross-domain operations — one sentence, and the car pulls over and parks nearby.
Smart driving is even more direct. The new version carries the largest on-device model, with end-to-end response speed up 300%; Infini-VLA lifts effective time-sequence to 30 seconds — a longer "memory" that lets the car keep deciding from what happened earlier. X-Foresight projects the next 6 seconds of surrounding vehicles, pedestrians, and environment.
In March, 2nd-gen VLA began mass delivery, with assisted-driving mileage over 50% in the first month. Next, XPeng will bring some Robotaxi L4-level experience to production cars; the G9L launches the new version first.
Parameters sound strong, but what decides the experience is whether it sees accurately, thinks clearly, and acts naturally in complex traffic. We'll see how this "remembers, predicts, thinks" VLA actually performs.
Figures
Source: Garry_高岩 (Weibo) — translated & adapted for XPENGTech.
Get the next one direct.
New articles, or a daily / weekly digest — the same list EVConduit keeps.