In a time when the narrative around artificial intelligence seems to have dwindled into monotony, a new player named DeepSeek has emerged, boldly disrupting the intelligent driving industryOn January 20, 2024, DeepSeek unveiled its DeepSeek-R1 model, simultaneously making the model's weights open sourceThe compelling combination of a low cost and extraordinary performance led to an astonishing user growth exceeding 100 million within just one week, establishing it as a global sensation.

In the realm of multimodal capabilities, DeepSeek further solidified its position by announcing the release of DeepSeek-VL2 in December 2024—a hybrid vision-language model aimed at facilitating advanced multimodal understandingThis model enhances its proficiency in visual question answering, optical character recognition, document, table, and chart comprehension, as well as visual localization, refining its utility across various applications.

Interestingly, before DeepSeek captured headlines, attention from the industry had already started drifting toward its modelsAutomaker companies began to experiment with the potential applications of DeepSeek's technology, igniting interest in what it could contribute to their efforts.

One high-ranking executive from a leading intelligent driving firm remarked, "The results have exceeded our expectations! Early this year, we conducted internal tests, and the most noticeable change with DeepSeek was the substantial improvement in model training efficiencyIt has reduced inference response times by 40% to 50%, all while optimizing resource utilization.” This feedback underscores the significant strides DeepSeek has achieved in enhancing operational protocols within intelligent driving.

Within the industry, experts harbor a collective belief that DeepSeek, as an open-source foundational model, holds the potential to expedite the training velocity of intelligent driving applications while simultaneously reducing associated costs

Advertisements

This positions DeepSeek as an essential tool for automation endeavors.

Further insights from Zhang Wei, the systems director at Zhixing Automotive Technology, delve into the intricacies surrounding the development of intelligent driving systemsHe notes that complex urban environments represent one of the most challenging domains for developersTraditional perception models struggle significantly when tasked with addressing long-tail scenariosCompanies are diligently working on developing visual language models (VLMs) to enhance their systems’ detection and processing capabilities for these complex situations.

However, the development of such advanced systems hinges on massive cloud computing resources and significant data training investmentsMoreover, models deployed within vehicles require robust hardware platforms with considerable processing powerHere, DeepSeek's distinctive technological advantages—such as its Mixture of Experts (MoE) architecture, Group Relative Policy Optimization (GRPO), and Multi-Head Latent Attention (MLA) mechanism—empower intelligent driving systems far more effectively.

Zhang succinctly summarizes the sentiment by asserting, "In a nutshell, DeepSeek allows urban autonomous driving functionalities to be accomplished with equivalent performance using fewer data and reduced training costs.” This encapsulates the breakthrough efficiencies that DeepSeek potentially brings to the field.

At the vehicle level, DeepSeek employs knowledge distillation to enhance model capabilities whilst reducing computational resource requirements and deployment costs on the vehicle's end

Advertisements

Advertisements

Advertisements