XPeng's Second-Gen VLA: The DeepSeek Moment for Chinese Autonomous Driving

XPeng Inc. is making an aggressive push to leapfrog Level 3 and achieve fully autonomous Level 4 driving, signaling a significant strategy shift for the Chinese electric vehicle maker. The company has unveiled a new end-to-end AI architecture it claims will be a watershed moment for the industry, backed by an annual investment of billions of yuan as it repositions itself as a leader in "physical AI."

At a technology event in Guangzhou on March 2, Chairman and CEO He Xiaopeng announced the company's second-generation Vision Language Action (VLA) model will be rolled out to customers via an over-the-air update in March 2026. He described the upgrade as a “DeepSeek moment” for autonomous driving—a reference to a breakthrough in AI that delivers a tangible, emergent leap in capability immediately apparent to everyday users.

The core of the new system is a completely overhauled end-to-end architecture that removes the intermediate language processing layer, allowing the model to directly translate raw visual data into driving commands. “A good smart-driving system in the future should be like taking an elevator,” He Xiaopeng said at the event, explaining that the goal is a system so intuitive and reliable that even his mother could use it without hesitation.

This technological pivot is accompanied by a bold strategic vision. He Xiaopeng is publicly calling on Chinese lawmakers to create policies that accelerate the development of L4 technology, bypassing the intermediate L3 stage. He predicts fully autonomous driving could be a reality within "one to three years," a timeline that underscores XPeng's ambition to evolve from an automaker into a foundational "embodied intelligence" company.

A New End-to-End Architecture

The most significant technical shift in XPeng’s second-generation VLA is the adoption of an end-to-end model that operates directly from visual input to action output. This eliminates the "L" (Language) as an intermediary, a layer that the company found created inefficiencies and limited the model's scaling potential.

According to Liu Xianming, the head of autonomous driving at XPeng, this breakthrough was enabled by several key innovations. The firm developed a native multimodal tokenizer capable of processing the continuous, unstructured data of the physical world, which is far more complex than text.

Furthermore, XPeng implemented a vision-based "Chain-of-Thought" (CoT) process. By preserving more details from visual information without the lossy conversion to and from a language-based model, this system allows for denser and more complex visual reasoning. Liu stated this new approach has improved efficiency by 32 times and reduced prediction errors by 33%. The system is designed to handle all scenarios, from narrow alleyways to unpaved rural roads, and even exhibit human-like emergent behaviors such as yielding to ambulances or responding to a traffic police officer's hand gestures.

Billions Invested in a "Turing" Compute Platform

This architectural overhaul is backed by a massive investment in compute power. He Xiaopeng disclosed to media that the annual investment in the foundational model runs into the "several billion yuan (hundreds of millions of U.S. dollars)" range. Liu Xianming noted that a single training run consumes around 50 petabytes of high-quality cloud data and involves approximately 4 trillion tokens, a scale comparable to training a large language model like ChatGPT.

To manage this immense data flow, XPeng developed its own AI compiler and a bespoke model architecture it calls "Turing." This full-stack optimization, from底层芯片 (underlying chip) to the model, allows the system to achieve dramatic efficiency gains. According to Liu, the enhanced "Turing" platform allows a single chip to deliver the effective compute power of ten Nvidia OrinX chips.

This enhanced efficiency has enabled an unprecedented pace of iteration. Liu said that over the past 13 months, XPeng increased single GPU training efficiency by 1,010% and single-task efficiency by 4,360%. This allowed the team to update its model 468 times in less than four months, equivalent to nearly four new versions per day.

Skipping L3, Pushing for L4 Regulations

XPeng is positioning its new system as "the first intelligent assisted driving system oriented towards L4." He Xiaopeng has made it clear that the company intends to skip L3 conditional autonomy and focus all its efforts on achieving L4, where the vehicle is fully responsible for driving in defined areas. He optimistically projects this could be achieved in as little as one to three years.

To accelerate this transition, He plans to submit a proposal during China’s upcoming "Two Sessions" parliamentary meetings. The proposal, titled "Suggestions on Accelerating the Transition of Autonomous Driving Technology from L2 to L4 and Improving Regulations and Management Policies," aims to lobby for a regulatory framework that supports the deployment of L4 technology.

He cited international developments to underscore the urgency, pointing to a "very aggressive" United Nations regulation allowing for hands-off driving in 2026 and a U.S. initiative to deploy 90,000 vehicles without steering wheels.

From Carmaker to "Embodied Intelligence" Firm

XPeng's announcement signals a fundamental re-imagining of its corporate identity. In an internal letter, He Xiaopeng has begun describing the company as an "embodied intelligence" firm, where autonomous driving technology serves as a foundational "physical AI" platform with applications extending to robotics and other intelligent carriers.

This vision involves a deeper fusion of autonomous driving with the in-cabin experience. The company has already merged its smart cabin and autonomous driving divisions into a single "General Intelligence Center" under Liu Xianming's leadership, anticipating that L4 capabilities will fundamentally change how users interact with their vehicles.

Once the second-generation VLA is widely deployed, Liu predicts that the daily token consumption from XPeng's fleet could reach 80 times the current total daily AI token calls in all of China. This highlights the immense data and intelligence requirements of physical AI, positioning XPeng not just as a consumer of AI, but as one of its largest operators and developers in the real world.