When the narrative of "AI 'on the car' has been flat, or even nearly flooded, DeepSeek has emerged as a strong force to" stir "the intelligent driving industry.

On January 20, DeepSeek officially released the Deepseek-R1 model and synchronized the open source model weights. With its low cost price and unusually strong performance, DeepSeek has achieved user growth of over 100 million in just 7 days, becoming a global phenomenon.

On the multimodal front, DeepSeek released its expert hybrid visual language model for advanced multimodal understanding, DeepSEEK-VL2, in December 2024, further enhancing the model's capabilities in visual question-answering, optical character recognition, document/table/chart understanding, and visual positioning.

In fact, when DeepSeek was unknown, the industry had already paid attention to its model, and some autonomous driving companies had begun to explore the application potential of the model in advance.

"The effect exceeded expectations! We actually tested it internally earlier this year, and the most intuitive changes we saw with DeepSeek were improvements in model training efficiency, which reduced inference response times by 40% or even 50%, while also reducing power utilization." An executive of a leading intelligent driving company revealed that.

In the industry's view, DeepSeek, as an open source basic model, is expected to accelerate the training speed of intelligent driving, reduce the training cost of intelligent driving, and become an important tool for intelligent driving training.

In terms of the development of intelligent driving systems, Zhang Wei, director of Zhixing Automobile technology system, analyzed that complex urban scenes are the most difficult place to develop at present, and it is difficult to rely on traditional perception models to solve such long-tail scenes. Various enterprises are trying to develop and train VLM models (visual language models) to optimize the system's detection and processing capabilities for long-tail scenes.

However, the development of such a system relies on a great deal of cloud computing power and data training costs, and the model deployed to the vehicle also relies on a hardware platform with a large computing power. DeepSeek through its unique technical advantages, such as MoE (hybrid expert architecture), GRPO (group relative strategy optimization), MLA (multi-head potential attention mechanism), etc., can better enable intelligent driving system development.

"In short, DeepSeek helps enable urban autonomous driving with the same performance with less data and training costs." Zhang Wei said.

Specifically, in the cloud training process, the data used for training the autonomous driving model can only be trained after being marked, and finally a deep learning model that can identify vehicles and pedestrians can be obtained. DeepSeek itself reduces the need for data annotation, so it can help intelligent driving companies with data mining and generation, reducing the cost of data collection and annotation.

On the vehicle side, DeepSeek can enhance model capabilities through distillation, reduce on-vehicle computing resource requirements, and reduce on-vehicle deployment costs. The computing power requirement and training cost of a single invocation of the model are greatly reduced.

In terms of scene understanding, Zhang Wei believes that DeepSeek has stronger logic and scene understanding after cross-modal migration, and its performance in extreme road conditions (such as broken roads, rare traffic sign recognition, sudden road construction, etc.) is expected to be better than the traditional model. Yuxin Yang, CMO (Chief Marketing officer) of Black Sesame Intelligent, also said that in the future, DeepSeek can be used to integrate multi-dimensional data such as vision, voice, and environment to achieve more anthropomorphic driving decisions, such as dynamically adjusting path planning at complex intersections, or quickly generating safety policies in unexpected situations.

Yang Yuxin believes that the core value of large models such as DeepSeek is to promote the upgrade of intelligent driving systems from "perception-driven" to "cognitively driven" through efficient end-to-side reasoning ability. If DeepSeek can achieve large-scale application through low-cost computing chips, it will accelerate the penetration of intelligent driving functions to the mass market. (Reporter Sun Xiaocheng Li Xingcai)