He Xiaopeng successively unveiled four major cutting-edge technologies. The humanoid robot took the stage in a catwalk. The L4-level vehicles that users can purchase are here.

He Xiaopeng: Three unmanned taxi Robotaxi models will be launched successively next year, and L4 cars with drivers are also coming!

According to a report from CheDongXi on November 5th, just now, He Xiaopeng officially announced four important technological achievements at the 2025 XPeng Technology Day.

The second - generation VLA large - model has received a significant upgrade and is open - sourced. Volkswagen has become the first strategic partner, and XPeng's Turing AI chip has also won a project from Volkswagen.

In 2026, XPeng will launch three Robotaxi models and start trial operations. It will also apply the entire set of L4 hardware and software for unmanned taxis to household models and launch Robo models for personal use.

XPeng's new - generation humanoid robot also stepped onto the stage with a "catwalk". The product will be mass - produced by the end of 2026, and He Xiaopeng will personally supervise the production.

▲ The new - generation IRON of XPeng makes its debut

The world's first mass - produced flying car, the "Land Aircraft Carrier", is expected to be delivered on a large scale in 2026.

At the end of the press conference, as an additional surprise, He Xiaopeng previewed the launch of the XPeng X9 Kunpeng Super Extended - Range Edition on November 6th.

This year's Technology Day is described by He Xiaopeng as "the most important one". One can sense that He Xiaopeng is in a very cheerful mood and can hardly contain his excitement.

▲ He Xiaopeng, Chairman and CEO of XPeng Motors

Since the first Technology Day in 2019, XPeng Motors has repeatedly turned technological concepts into reality. This time, XPeng clearly presented the complete strategy of "Physical AI" to users.

In the next decade, with a firm look in his eyes, He Xiaopeng aims to be a pioneer in mobility in the Physical AI world and build a globally - oriented embodied intelligence company.

01. The second - generation VLA receives a significant upgrade and will start being pushed in the first quarter of next year

At the beginning of the press conference, He Xiaopeng clarified the theme of this event.

▲ He Xiaopeng clarifies the theme of the press conference

He Xiaopeng believes that the integration of the digital and physical worlds will give rise to Physical AI, which means that machines will gradually gain the ability to understand, interact with, and change the world.

▲ He Xiaopeng explains Physical AI

The first change brought by Physical AI is the operating system, and the first core upgrade at the XPeng Technology Day is the advanced driver - assistance system - the launch of the second - generation VLA.

He introduced, "In 2024, XPeng began exploring the VLA solution. One is the standard V - L - A base model, and the other is the V - A base model, led by Liu Xianming (Head of the Autonomous Driving Center at XPeng Motors)."

▲ The two VLA system architectures of XPeng

After investing 30,000 GPU cards in computing power and spending over 2 billion yuan on training costs, XPeng finally developed a new, safer, and smoother autonomous driving system - the second - generation VLA of XPeng.

▲ The iterative process of XPeng's advanced driver - assistance system

He Xiaopeng said that XPeng's second - generation VLA has pioneered a new paradigm for the physical world model. It is the biggest leap in intelligent driving after end - to - end and standard VLA.

▲ XPeng's second - generation VLA

This system eliminates language translation and takes vision as the core. It can reason about the world like a human, with higher reasoning efficiency. It can be directly trained using nearly 100 million real - world video data, equivalent to the sum of extreme driving scenarios encountered in 65,000 years of driving. It can predict future scenarios through visual reasoning and generate training scenarios based on real - world simulations.

▲ The architecture of XPeng's second - generation VLA

How does the second - generation VLA achieve these?

▲ The architecture of XPeng's second - generation VLA

This system uses a 30,000 - GPU - card autonomous driving cloud intelligent computing cluster and a cloud - based large base model with 72 billion parameters. The model is iterated across the entire chain every five days.

▲ The computing power resources of XPeng's second - generation VLA

XPeng has also promoted the optimization and innovation of the entire chain from chips to operators to models. On the Ultra models, it has achieved the industry's highest effective computing power of 2250 TOPS at the vehicle end. By optimizing operators specifically, the model reasoning efficiency has been increased by 12 times. Based on the collaborative optimization of hardware and software, a model compiler and software stack have been developed specifically, and the model parameters have been increased by 10 times compared to the industry mainstream.

▲ The full - chain capabilities of XPeng's second - generation VLA

In terms of intelligent driving performance, the second - generation VLA has shown amazing results. The Small Road NGP function is about to be launched, which can smoothly cover highways, cities, small roads, and industrial parks. The average takeover mileage on complex small roads has increased by 13 times.

▲ The "Small Road NGP" function of XPeng's second - generation VLA

XPeng also compared it with Tesla's FSD V13. On a 20 - kilometer complex small road, XPeng's second - generation VLA only needed to be taken over once, while FSD needed to be taken over five times.

▲ The comparison between XPeng's second - generation VLA and Tesla's FSD

The second - generation VLA also has a new intelligent emergence ability and the generalization ability for all - weather and all - scenario situations, such as recognizing traffic police gestures and predicting traffic light logic in advance. He Xiaopeng commented that it was "like discovering a new continent".

▲ The stop - on - signal function of XPeng's second - generation VLA

The second - generation VLA also makes its debut with the non - navigated advanced driver - assistance (Super LCC+) human - machine co - driving function, which can be activated anywhere as long as it complies with global laws and regulations.

▲ XPeng's second - generation VLA will debut the "non - navigated advanced driver - assistance" function

In terms of the push schedule, XPeng will open media test drives this month, offer a co - creation experience for pioneer users by the end of December, and fully push it to Ultra users in the first quarter of 2026.

▲ The push time of XPeng's second - generation VLA

He Xiaopeng also announced that the second - generation VLA will be open - sourced to global business partners. Volkswagen Group will be the first customer for XPeng's second - generation VLA, and XPeng's Turing automotive chip has also won a project from Volkswagen.

▲ Volkswagen will be the first to use XPeng's second - generation VLA

After the linkage between XPeng's second - generation VLA and VLM, new capabilities will emerge: ultra - low - latency interaction with the autonomous driving system and an AI butler that can seamlessly switch between multiple languages.