XPeng CEO He Xiaopeng announced at the G9L launch that the second-generation VLA dual Orin-X distillation version will roll out in Q4 2026, with the single Turing chip version arriving in late August to early September. This upgrade means Max models launched as early as 2022 will still receive cutting-edge autonomous driving capabilities.
Second-Generation VLA: The Core Upgrade of XPeng's AD Stack
On August 12, 2026, at the XPeng G9L launch event, He Xiaopeng unveiled the roadmap for the second-generation VLA (Vision-Language-Action) assisted driving upgrade. According to the plan, the distillation version based on a single Turing chip will be released at the end of August to early September; the dual Orin-X chip version, whose performance "exceeded expectations," is scheduled for Q4 2026.
This marks another major iteration since XPeng's first batch of second-generation VLA pushes in March 2026. The core goal is to bring near-L4 autonomous driving experience through the synergy of large models, massive compute, and big data.
Distillation Tech Explained: Balancing Performance and Compute
The distillation approach compresses the full Ultra model's core capabilities onto lower-compute hardware platforms. He Xiaopeng previously clarified the differentiation:
- Max Version: Industry-leading L2 assisted driving, covering highways and main urban roads
- Ultra Version: Built for L4 capability, supporting full-scenario navigation
A comparison of hardware platforms reveals significant compute differences:
| Hardware Platform | Effective Compute | Representative Models | Rollout Schedule |
|---|---|---|---|
| Single Turing Chip | ~750 TOPS | 2026 X9 etc. | Late Aug - Early Sep 2026 |
| Dual Orin-X | ~508 TOPS (2×254) | 2022 X9 etc. Legacy Max | Q4 2026 |
| Dual Turing Chips | Up to 2250 TOPS | P7 Ultra etc. | Rolled out March 2026 |
He Xiaopeng emphasized that the distillation version retains over 90% of the full model's safety and smoothness, with confidence to surpass all industry L2 systems. Early testers reported that the distillation version "fully preserves the VLA's signature safety and fluidity," with travel efficiency improved by 23%.
Legacy Owner Care: Orin-X Models Keep Evolving
The most notable aspect of this upgrade is XPeng's commitment to legacy Max owners. Dual Orin-X Max models launched in 2022, now nearly four years old, will still receive second-generation VLA capabilities. He Xiaopeng repeatedly stressed "we will definitely take care of our owners" in live streams.
This reflects an industry trend: smart driving has shifted from "hardware arms race" to "software-defined capability." Through model compression techniques like distillation, automakers can extend the lifecycle of older hardware. For consumers in Central Asia, Russia, and emerging markets, this means their Chinese EVs can stay "ever-new," significantly enhancing total cost of ownership.
Competitive Landscape: VLA Becomes Industry Consensus
In 2026, VLA large models have become the shared choice among China's leading EV startups. XPeng and Li Auto have each released their VLA solutions—XPeng's second-gen VLA and Li Auto's MindVLA-o1—with distinct approaches:
- XPeng: Pure vision + VLA large model, self-developed Turing AI chip for hardware-software integration, compute utilization boosted from industry average 22.5% to 82.5%
- Li Auto: VLA + LiDAR fusion, deep family-user integration, voice driving command recognition accuracy at 98%
XPeng's second-generation VLA delivers impressive metrics: 23% overall driving efficiency gain, 76% narrow-road efficiency gain, 69% mixed-traffic efficiency gain, 28% rush-hour efficiency gain. The company expects full autonomy within 1-3 years and plans to launch Robotaxi operations by end of 2026. For more AD tech updates, visit EX1000.COM.
