The assisted driving models are getting larger and larger. XPeng and Li Auto have first entered the order of magnitude of 7 billion parameters.

The assisted driving systems of automobile manufacturers are accelerating their shift towards AI. A clear indication is that the parameters of the in - vehicle assisted driving models of leading new - force automakers are approaching the parameter magnitudes of many large AI models.

36Kr Auto has learned that the large model that XPeng Motors is about to deploy on the vehicle will have at least 7 billion parameters. Another leading new - force automaker, Li Auto, will also have its in - vehicle large model parameters reach the level of 7 billion after its self - developed assisted driving chip is installed in vehicles next year.

This number of parameters is close to the general parameter magnitude of large AI models.

AI Layouts of XPeng and Li Auto

XPeng's in - vehicle large model is distilled from the cloud - based large model "XPeng World Base Large Model" that is being developed internally. The main reason for this is to address the problem that the computing power, storage, and memory bandwidth of in - vehicle assisted driving chips are insufficient, making it impossible to directly deploy large models on the vehicle.

In the second half of 2024, XPeng Motors began to move towards cloud - based large models. Currently, XPeng is developing an ultra - large - scale autonomous driving large model with a base of 72 billion parameters, the "XPeng World Base Model", which will be released at next month's AI Technology Day.

At the AI Technology Sharing Conference in April this year, XPeng introduced that this cloud - based large model uses LLM as the backbone network and is trained with a large amount of multi - modal driving data. It has the capabilities of visual understanding, chained reasoning, and action generation. After XPeng completes the training of this model in the cloud, it will "extract the essence" and deploy the distilled small model on the vehicle.

This method refers to the knowledge distillation route already used by DeepSeek, which is essentially model compression.

To ensure that the distilled large model can be successfully installed in the vehicle, XPeng has made a series of preparations at the hardware and R & D resource levels.

At the hardware level, since 2020, XPeng has initiated the self - development of the "Turing" AI assisted driving chip. In June this year, this chip was officially mass - produced and launched, and was first installed in the 2025 XPeng G7.

It is a chip specially designed by XPeng for AI requirements and end - to - end large models. Its AI computing power is approximately 700 Tops, close to that of NVIDIA's latest AI chip Thor, and it can handle large models with up to 30 billion parameters at most.

In addition to the hardware preparations, at the beginning of August this year, XPeng Motors held a mobilization meeting for its autonomous driving center, which was personally presided over by He Xiaopeng. At the meeting, He Xiaopeng proposed to allocate all AI resources to the base model team to support the installation of this 7 - billion - parameter World Base Model in the vehicle.

Li Auto, which achieved the installation of a large model in the vehicle earlier, has also not relaxed its pursuit of the AI wave.

Li Xiang, the CEO of Li Auto, said at the second - quarter earnings conference call this year that currently, the number of parameters of Li Auto's in - vehicle large model is more than 4 billion, more than 10 times that of the previous end - to - end model. 36Kr Auto has learned from multiple industry insiders that next year, after Li Auto's self - developed assisted driving chip is installed in the vehicle, the parameters of the VLA model deployed on the vehicle by Li Auto will also reach more than 7 billion.

Initially, Li Auto installed a small VLM large model with fewer parameters and slower operation speed in the vehicle to achieve the installation of a large model.

In October last year, Li Auto launched an assisted driving solution based on end - to - end + VLM. In this solution, the end - to - end system is the fast system, and the VLM is the slow system, and the two systems work simultaneously.

This solution is deployed on a dual - Orin X controller, with each Orin serving a model independently.

In this context, the end - to - end system is like the brain sitting in the driver's seat responsible for driving, while the VLM model can only play the role of sitting in the passenger seat and occasionally helping to watch the road, making it difficult to fully exert the strength of the large model.

However, compared with the end - to - end system, Li Auto currently prefers to embrace the VLA model. As of now, Li Auto has pushed the VLA driver large model to all AD MAX vehicle users.

The VLA model was first launched by Google AI company Deepmind and has since become the mainstream technical paradigm and framework in the field of embodied intelligence. Since the VLA model has a complete brain system, with language and thinking - chain reasoning abilities, it can both see and understand and truly execute actions, which conforms to the way humans operate. Therefore, this technology is now also applied in the assisted driving field by automakers such as Li Auto and XPeng.

To promote the installation of a VLA model with larger parameters in the vehicle next year, Li Auto made significant adjustments to its organizational structure this year: In May this year, Xia Zhongpu, the former person in charge of the end - to - end system at Li Auto, left the company. Last month, Li Auto split its assisted driving team into 11 secondary departments to promote the R & D of large AI models with a flatter organization.

In addition to XPeng and Li Auto, Huawei's WEWA architecture adjusts the in - vehicle world model through the cloud - based world engine, and NIO has also deployed a large world model in the vehicle.

AI seems to be defining the assisted driving of automakers.

Large AI Models ≠ Better Assisted Driving Performance

On the other hand, Tesla, which is currently recognized in the industry as the furthest - advanced in the field of assisted driving, has achieved regional Robotaxi with end - to - end technology. Assisted driving suppliers such as Horizon and Momenta have also achieved good assisted driving performance through end - to - end technology.

In contrast, the assisted driving performance of some automakers obsessed with the AI narrative has been quickly caught up, and even the vehicle control experience of some models has been surpassed.

That is to say, companies such as Tesla and Momenta have achieved better results with fewer in - vehicle model parameters.

In a sense, this may indicate that there is no inevitable relationship between the parameter magnitude of AI models and the assisted driving effect.

End - to - end technology emphasizes imitating human driving behavior, while the advantage of large models lies in logical reasoning ability, emphasizing thinking like a human. The core of assisted driving technology is spatial perception, and the logical reasoning ability of large models is only used in a few scenarios.

Therefore, if automakers blindly install models with larger parameters without improving the end - to - end experience, it means that most of the in - vehicle computing power resources will be tilted towards the reasoning process of large - language models, leaving only a small amount of resources for spatial perception, which will instead lead to a regression in the assisted driving experience.

So, what is the driving force behind automakers' eagerness for large models?

On the one hand, the scope of automakers is gradually expanding. Some automakers not only want to build cars but also want to engage in embodied intelligence. A typical example is Li Auto.

At the live - streaming room of the Li Auto AI Talk event at the end of last year, Li Xiang himself redefined Li Auto as an artificial intelligence company. Although it will continue to build cars, it will regard cars as spatial robots in the era of artificial intelligence and use cars as an application scenario for its artificial intelligence.

And the VLA that this company currently advocates is precisely the mainstream technical paradigm and framework in the field of embodied intelligence.

Another example is XPeng. It is not only a member of the VLA camp, but according to its plan, the Turing AI chip will not only be installed in cars in the future but also used in AI robots and flying cars. This shows that XPeng intends to transfer the capabilities accumulated in assisted driving to flying cars and embodied intelligence.

On the other hand, automakers' pursuit and promotion of large AI models may also have some marketing purposes.

Recently, the performance of Chat - GPT has given large AI models a huge cross - circle effect. This is similar to the situation after Tesla V12 was widely pushed in North America, which once made end - to - end a hot marketing term in the field of assisted driving.

Now, automakers are moving large AI models into in - vehicle assisted driving systems. Although there are technical considerations behind this, to some extent, this may also be a marketing method for companies to seize the high ground of public opinion.

However, regardless of the purpose, improving the assisted driving experience should be the top priority for automakers at present. An AI narrative that deviates from this goal is somewhat like putting the cart before the horse. Improving the end - to - end assisted driving experience may be the best way for automakers to correct their course at present.

Author's WeChat: luckg17305264638