Phoenix.com Technology News (Author/Yu Lei) On April 14, Xiaopeng Motors held an AI technology sharing session in Hong Kong today, officially disclosing its 72 billion parameter ultra-large-scale autonomous driving model - the "Xiaopeng World Base Model". This model will not only provide Xiaopeng Motors with a brand new intelligent driving "brain", but will also be deployed to the vehicle terminal through cloud distillation technology, while empowering Xiaopeng's AI robots, flying cars and other terminal devices.

According to Li Liyun, the head of Xiaopeng Automobile's autonomous driving, the base model is based on a large language model as the backbone network and is trained through massive high-quality driving data. It has three core capabilities: visual understanding, chain reasoning and action generation. Through reinforcement learning training, this base model can continuously evolve itself and is expected to develop autonomous driving skills comparable to or even surpass humans.

The parameter scale of the Xiaopeng World Base Model is as high as 72 billion , about 35 times that of the mainstream VLA model. Its significant advantage lies in its ability to have chain reasoning (CoT), which can perform complex common sense reasoning like humans, and convert the reasoning results into control signals such as steering wheel and brakes, so as to achieve effective interaction with the physical world.

In order to support the research and development of base models, Xiaopeng Motors has built a "cloud model factory" from scratch and built the first Wanka smart computing cluster in the domestic automobile industry. At present, the "factory" has 10 EFLOPS of computing power, and the cluster operation efficiency remains above 90% all year round. The full-link iteration cycle from cloud to end can reach an average of once every 5 days. The data infrastructure independently developed by Xiaopeng Motors has increased the data upload scale by 22 times, the data bandwidth during training has increased by 15 times, and the model training speed has increased by 5 times.

At the sharing meeting, Xiaopeng announced three phased results: it verified that the scale law continues to take effect in the field of autonomous driving ; it successfully realizes the basic model control vehicle on the vehicle end with the after-installation of computing power; it starts the 72B parameter basic model training to build a model training framework for reinforcement learning.

Through a large number of experiments, the Xiaopeng team verified the applicability of the "Law of Scale" in the field of autonomous driving for the first time, that is, the larger the parameter scale and the more training data, the stronger the model's ability. Based on the experience accumulation of past "rule era", the R&D team also designed an effective reward function for reinforcement learning, transforming rule experience into the productivity of training base models.

At the same time, Xiaopeng Motors has begun to develop the World Model, which is an important part of the "cloud model factory" and supports the performance optimization of the base model. This world model can model and feedback in real time, simulate the real environment state based on action signals, and build a closed-loop feedback network to help the base model evolve continuously.

He Xiaopeng, chairman of Xiaopeng Motors, said: " Our goal is to become the number one in the world of physics and promote huge changes in the fields of automobiles, robots, and flying cars. " It is reported that Xiaopeng Motors will further share the details of the research and development and training of the Xiaopeng World dock model in June this year at the CVPR of the International Top Conference on Computer Vision.