From February 21st to 23rd, the 2025 Global Developer Pioneer Conference was held in Shanghai.

Transwarp (688031. Yang Yifan, vice president and director of AI R&D, told The Paper during the conference that DeepSeek is "proofing" AI open source, and open source is the trend. For practitioners in the artificial intelligence industry, open source superimposed on the "moat" of the enterprise's own business, combined with the understanding of scenarios and data, is a hot spot for future development. DeepSeek lowers the threshold for application, and the upper limit of application will be higher in the future.

Lu Yang, assistant to the chairman of Beijing Jiuzhang Yunji Technology Co., Ltd., also told The Paper during the conference that software is convergent, the threshold for AI application development is not in the software layer, and privatized data and industry know-how (industry knowledge) are the core competitiveness.

Lu Yang said that after the emergence of DeepSeek, the essence of large models empowering vertical applications has not changed, and computing power is applicable to a wider range of industries. In the past, computing power empowered industries such as intelligent driving, robotics, and biomedicine, but now there is a demand for computing power in all walks of life, "Some blind date companies have also approached us to buy computing power, train a blind date model, analyze it with a large model, match the information of both parties online, and customer feedback is more accurate." ”

Talking about the industry impact brought by DeepSeek, Yang Yifan said that the use of large models is not only the inference model itself, but also includes the ecological construction of large models, the management of large model computing power, the construction of knowledge base system, and the development of agent applications. Enterprise-level business logic is complex and has high stability requirements, and DeepSeek promotes the implementation of inference models on the production side. He said that the full-blood version of DeepSeek is a 671B FP8 precision model, and different distillation versions are also produced, and enterprises can choose different versions of the model to handle different complex tasks according to business needs. The inference model works with the small model to form workflows, data flows, and operation flows, and has a deeper understanding of the task, which has higher requirements for AI infrastructure.

At present, Transwarp's web version and applet are connected to the full-blooded version of the DeepSeek 671B inference model, and users can open the application by turning on "Deep Thinking". Enterprise users can get the advanced version of the DeepSeek solution, which supports the rapid deployment of DeepSeek large models in enterprise privatization environments, helping enterprises quickly develop DeepSeek-based internal applications. Transwarp's "Boundless AI PC Edition" is connected to DeepSeek to realize the localized operation of the DeepSeek large model on personal computers. The AI PC supports "cloud collaboration" and is equipped with a local enhanced retrieval system, which breaks through the limitations of traditional keyword search, supports deep semantic understanding, and supports accurate knowledge base retrieval through retrieval enhancement generation. The model inference is all completed locally, and the private data is zero out of the network, ensuring data security and controllability.

Jozzon Yunji also launched the DeepSeek large model all-in-one machine, which is ready to use out of the box and is equipped with 16 high-performance GPU cards. Lu Yang said that DeepSeek has brought about a change in the model architecture, and the previous pre-training superposition and fine-tuning and inference mode has become post-training, and the model will also reverse fine-tune the model in the process of continuous answer or inference, so that the combination of training and inference and the demand for computing power is increasing.

Yang Yifan believes that the reasoning ability of traditional agents is relatively weak, and the inference model represented by DeepSeek can help give birth to more business formats such as knowledge base and agent construction. The current development trend of AI also confirms that AI talents cannot be limited to understanding algorithms, and compound AI talents are needed in the era of large models, who must have an understanding of the whole link such as corpus construction, model training, model implementation, engineering acceleration, and enabling production, and also specialize in one field.