Over 20 car companies collectively stand in line, DeepSeek is the savior?

At the beginning of 2025, the automotive industry was stirred up by a large AI model called DeepSeek. From Geely and BYD to Dongfeng and Changan, more than 20 car companies are competing to integrate this technology, and even car infotainment service providers such as Zebra Intelligence and Yikatong are following suit. It seems that a new car without the DeepSeek label is difficult to even proceed

 

At the height of this intelligent arms race, even more shocking news is coming from across the ocean. On February 18th, xAI, an artificial intelligence company under Musk, released the Grok 3 models. This tech giant claims that the Grok 3 is "the smartest artificial intelligence on Earth" and has announced that it will be fully integrated into Tesla's entire lineup.

 

This heavyweight news has pushed the global automotive industry's intelligent competition to a dramatic climax. However, behind the market frenzy, two important questions have also begun to emerge: how much investment can be converted into real user experience when the capital market is celebrating AI concepts with real money and silver? Will DeepSeek accelerate the reshuffling of the automotive industry?

 

A dual battle between technological breakthrough and capital game

 

One of the significant advantages of DeepSeek over long established OpenAI models is its lower training cost, which is only 5% to 10% of similar OpenAI models. For example, the training cost of DeepSeeker R1 is $6 million, while the estimated cost of GPT-4 is as high as several hundred million dollars.

In addition to the advantage of training costs, DeepSeek also plays a key role in reducing inference costs. The launch of open source models enables small and medium-sized enterprises and developers to participate in AI competition at a lower cost, thus strongly challenging the "moat" built by closed source vendors such as OpenAI.

 

At the same time, Musk is also racing against his competitors in terms of technology for the Grok 3. In multiple benchmark tests, the reasoning models Grok-3 Reasoning and Grok-3 mini Reasoning have significantly outperformed o3-mini (high), o1, DeepSeek R1, and Gemi-2 Flash Thinking in mathematics, science, and coding. However, Musk is not in a hurry to open source the Grok 3 and will only release the previous generation model Grok 2 for now.

 

 

However, Grok3 is expensive, and Musk spent 100000 H100 supercomputers to train the "Beast" product. On the other hand, DeepSeek has taken a low-cost, high-performance, more pragmatic improvement path, and most importantly, is open source.

Industry insiders believe that DeepSeek may perform better than Grok 3 in training and reasoning with Chinese language corpus, which is also one of the reasons why DeepSeek has sparked strong reactions in the automotive industry.

 

On the one hand, DeepSeek does provide many imaginative spaces for car companies. Especially, the extremely low inference cost of the DeepSeeker R1 model has become a huge attraction for traditional car companies. If these car companies choose to independently develop AI systems, they may need to invest billions of dollars and years of time, while accessing DeepSeek only requires a relatively small amount of cost to obtain powerful AI capabilities.

 

On the other hand, DeepSeek has recently attracted a lot of attention and capital support. On February 7th, Geely announced the integration of its Star Rui large model with DeepSeeker R1. With the support of intelligent concepts, Geely Automobile's stock price soared by 26% in the following five trading days.

 

In the increasingly competitive automotive market environment, car companies are paying more and more attention to intelligent marketing, and DeepSeek has become an indispensable hotspot.

However, the current collaboration with DeepSeek shows that its applications in the automotive field mainly focus on the interactive functions of intelligent cockpits. Many opinions in the industry believe that the practice of some car companies accessing DeepSeek is suspected of "riding on traffic", and cooperation is mostly limited to upgrading basic functions such as voice control, rather than truly integrating underlying technologies.

 

 

The technological reefs under the halo

At present, there are probably two modes of application for DeepSeeker R1 in the automotive field:

The first mode is for car companies to use DeepSeek's open source protocol to deploy the complete R1 model or optimized simplified version to their own servers. Enterprises can fine tune the model and combine it with their existing models to form a more powerful comprehensive model. In this way, car owners can experience the voice interaction function provided by this new model through the in car system networking.

 

The second mode is to compress the R1 model into a very small model and update it to the in vehicle device. In this way, even if the car owner does not have a network connection, they can still use some of DeepSeek's functions and enjoy the convenience of intelligent experience.

 

The main advantage of DeepSeek is that it provides every car company with the opportunity to access large language models, almost without incurring additional costs, and can effectively solve the problem of fuzzy semantic speech dialogue Zhang Ming (a pseudonym), an auto analyst, said, "But the current situation is that this model probably cannot directly call the on-board functions, which means that auto companies must develop their own real-time voice interfaces."

 

So, what practical perceptual experiences can users gain from DeepSeek's application?

 

Traditional in car voice assistants are often criticized for their artificial intelligence, while DeepSeek's semantic understanding ability makes the in car system more user-friendly, with smoother conversations and no longer mechanical and dull. For example, the cabin of the Lantu Zhiyin model can automatically adjust the air conditioning or seat heating according to the user's vague instructions (such as "I'm a bit cold"). This "human like interaction" experience hits the urgent need of users for "emotional intelligence". At the same time, they have a better understanding of cars and passengers, and are able to control vehicles more intelligently based on geographic location, weather information, user habits, etc., and provide suggestions for after-sales maintenance.

 

As long as car companies can provide practical and feasible solutions, there is still a lot of room for improvement in the in car experience. Perhaps this can increase the overall rating from the current 60 points to around 75 points, "Zhang Ming pointed out.

If you look at Tesla, although the car now has voice interaction function, it can basically only respond to simple commands. If based on the performance of Grok 3, the large model of getting on the car will be able to provide users with a more natural voice conversation service.

 

 

Although voice assistants have become more user-friendly, their intuitive effects still fall short in improving driving operations, active safety, and overall interaction experience.

This is also part of the reason why among the more than 20 cooperative car companies, the top new force brands are "inactive" or "slow to take action", because similar functions have already been implemented in their models.

 

For example, NIO's NomiGPT has achieved emotional dialogue capability through a personification design that combines software and hardware, and is regarded by users as the 'soul of NIO'; The voice assistant 'Ideal Student' under Ideal Auto has introduced its self-developed cognitive system MindGPT, which supports various functions such as voiceprint recognition and content recognition.

In terms of ecological construction, DeepSeek has not yet formed its own ecosystem.

 

For example, DeepSeek is unable to recommend nearby restaurants to users, while systems such as Huawei HarmonyOS and Xiaomi's Pengpai OS have already built a cross device user experience through the ecological collaboration of mobile phones and car devices, forming a certain differentiation barrier.

The first tier enterprises still have significant advantages in large models, computing power, and data accumulation, and they may think that DeepSeek's incremental value is limited.

 

An insider from NIO admitted, "Even if some companies sign an interface agreement and cooperate, NIO is still researching the potential application of DeepSeek in vehicles, but it is not so simple to implement, so it will not easily announce deep integration with DeepSeek now

 

 

Xiaopeng Motors Chairman and CEO He Xiaopeng also expressed a similar view during a live broadcast, "Currently, more than 20 domestic car companies have announced the integration of DeepSeek, many of which only call the interface through the web. This method is not of great value to the final driver, and it really needs to achieve excellent user experience

 

Can DeepSeek knock on the door to 'intelligent driving for all'?

 

When car companies limit DeepSeek to voice interaction and in car entertainment, the true value of this AI technology is being severely underestimated.

 

The Bank of America Securities report reveals that DeepSeek resonates deeply with the underlying logic of autonomous driving technology - its multimodal processing capabilities help them better understand road scenarios and provide more powerful performance in complex road conditions.

 

The wave of technological revolution is surging at an even faster pace. Users speculate that the reasoning and multimodal data processing capabilities of Grok 3 can also be integrated with Tesla FSD. For example, when Grok 3 is connected to Tesla's vehicle sensor network and real-time information flow, it can analyze camera and radar data in real time.

 

BYD Wang Chuanfu's strategic layout is more profound. When high-end intelligent driving breaks the price red line of 100000 yuan, this technological revolution is moving from the ivory tower of luxury cars to the general public, consistent with DeepSeek's direction of low cost, low computing power, and high performance.

 

 

He Xiaopeng believes that DeepSeek has two technical details that match Xiaopeng Motors' judgment: first, distillation technology is an effective method for preserving model capabilities; Secondly, the distillation effect of large models is stronger than that of small models in reinforcement learning.

 

The core goal of DeepSeek's "distillation method" is to extract knowledge from complex models into simple models. For example, previous large-scale model training was similar to a large number of question sea tactics, while distillation is a method of treating refined and excellent large-scale models as "teachers" of new models, screening out effective problems, and guiding the training of new large-scale models.

 

In addition, multimodal models can generate rare driving scenario data, helping to reduce the cost of actual road testing. Industry insiders have pointed out that adopting DeepSeek can save about 90% of annotation costs and generate a large amount of enhanced data. For example, by comparing the driving style data of experienced drivers, the trajectory generated by the model can be evaluated, which also helps to improve its performance in extreme road conditions (especially urban NOA).

 

Unlike traditional AI training methods, DeepSeek innovatively adopts a reinforcement learning training mode, which not only overcomes the dependence on manually annotated data, but also subverts the traditional logic of the "computing power arms race" in the past.

 

However, the road to the future is not smooth. From the current situation, few car companies dare to directly integrate DeepSeek into intelligent driving systems, and their understanding and development of DeepSeek are still in their early stages.

 

Intelligent driving requires extremely high safety, and the illusion problem of large models is an inherent flaw, and there is currently no perfect solution, "said an autonomous driving practitioner."

 

 

When using models, car companies must be cautious about the information they generate, and ensuring the safety and reliability of their functions requires a lot of training time. In addition, due to the differences in computing power and architecture among different vehicle models, even if distillation technology is used, targeted development still needs to be carried out according to specific vehicle models in the future.

 

Therefore, in order to directly apply DeepSeek to autonomous driving, there is still a lot of room for improvement in terms of technological depth and practical applications, and a long exploration process is needed.

 

As Huawei Terminal Chairman Yu Chengdong stated, advanced intelligent driving is completely different from being "usable" and "easy to use and safe"; Li Liyun, the person in charge of Xiaopeng's autonomous driving, has also said, "There will be many car companies that can ultimately have intelligent driving capabilities, but most of them may be around 60-70 points. Those who truly achieve 80-90% or more are definitely self-developed and truly oriented towards AI reform

 

The technological dividends brought by DeepSeek are real, but the true value reflected after its implementation is still unclear. However, it can be certain that with intelligence being re driven by AI, deeper reshuffling undercurrents have quietly emerged in the automotive industry.