GuShiio.com 
From February 21st to 23rd, the 2025 GDC Global Developer Pioneer Conference will be held in Xuhui, Shanghai. The theme of the conference is to shape the world’s infinite possibilities, focusing on companies in various scenarios such as finance, medical care, and education intelligent manufacturing, and technology providers from multiple fields such as AI development tools, multimodal large models, enterprise-level services, and open source ecosystems. Explore the latest directions for the industrialization of large models and the forefront of technology.
At the opening ceremony of the GDC on the 22nd, Shen Xiangyang, Chairman of the Board of Governors of the Hong Kong University of Science and Technology and a foreign academician of the National Academy of Engineering, Qi Yuan, President of the Shanghai Institute of Science and Intelligence and Haoqing Distinguished Professor of Fudan University, Jiang Daxin, founder and CEO of Step Star, and Yang Fan, co-founder and vice president of Shangtang Technology, shared and discussed topics such as AI big models, DeepSeek, open source big models and Agent (agent) applications.
On the afternoon of the 22nd,AI company Shangtang Technology also announced that it can build a multi-agent large-scale model application development framework LazyLLM by open source. Shangtang’s large device has launched the Vientiane platform, which includes a model development platform and an Agent application development platform. It has also launched the DeepSeek series of models. Enterprise customers and developers can get 10 million tokens for free use within 3 months in the Shangtang Large Device Vientiane platform; At the large-scale model application level, Shangtang released a new code Little Raccoon 2.0 and Office Little Raccoon 2.0, realizing the transformation from Copilot to Agent. Also,Wang Xiaogang, co-founder and chief scientist of Shangtang Technology, revealed that the New Riven-Day Model 6.0 will also be released this year.
Obviously, entering 2025, AI agent applications and open source model software will become a new trend.
Opening ceremony: Shen Xiangyang, Qi Yuan, Jiang Daxin and others talked about open source and the prospects of agents
In the past year, Shanghai has continued to increase AI industry policy innovation, promote technological research, and strengthen the ecological layout of the entire industrial chain.
Public information shows that in 2024, the scale of Shanghai City’s artificial intelligence industry will exceed 450 billion yuan, a year-on-year increase of more than 7%, and a total of 60 generative AI models have been registered.
According to the “Implementation Plan on Artificial Intelligence Molding Shenzhen City”, by the end of 2025, Shanghai will build a world-class artificial intelligence industrial ecosystem covering computing power, corpus, models, applications and other aspects, and create a world-class high-end AI industry cluster.
With the emergence of the domestic DeepSeek model, new AI scenarios and new applications continue to emerge. As one of the important gathering places for AI models and computing power industry chain companies, Shanghai and other places may enter the innovation stage of the AI cycle more quickly, promoting the formation of new industrial forms, spawning new employment scenarios, and potential consumption growth.
Harry Shum, a foreign academician of the National Academy of Engineering, fully affirmed the power of open source in the era of big models when delivering a keynote speech entitled “Innovation and Thinking in the Era of Big Models.” nbsp;
“DeepSeek shows everyone the victory of the open source community. By opening up the great model, more people will have the opportunity to do more great things on this model. rdquo; Shen Xiangyang pointed out that from the perspective of technological progress, open source is not new, but its vigorous development began after the emergence of the Internet. The Internet has greatly enhanced the ability to cooperate and open source globally. In addition, open source can also be understood from the perspective of business choices.
Shen Xiangyang believes that looking at the proportion of models developed by different companies in actual use by enterprises from 2023 to 2024, among large models with market shares ranging from high to low, the open and closed source strategies are different and have no obvious relationship with their market shares. Therefore, he believes that open source and closed source are not opposing business models. Open source is still an important area of global cooperation. China is a beneficiary of open source research and now a contributor to the international open source community.
Shen Xiangyang emphasized that at present, the share of closed-source models still far exceeds the share of open source, but this matter may change very much in the next year or two. Everything must have a balance, so we need to think about the future.
Shen Xiangyang said that for large models, the previous research focus was on the GPT model, but now the focus is on the Reasoner model, which is a new learning paradigm.
He pointed out that previously represented by OpenAI’s GPT series models, this is a fast thinking mode. Large models are all pre-training and expansion. The main principle is to predict the next token, while the reasoning model focuses on task completion and allows repeated trial and error. and correction, closer to human slow thinking. The model will make a draft first, and after repeated trial and error to find the correct path, the process and answers will be summarized to train the model’s ability to think slowly. Therefore,This new model has been applied in areas such as visual reasoning by gradually abstracting complex problems and deriving solutions.
Should the relationship between humans and machines, which has been widely discussed recently, be AI (Artificial Intelligence) or IA (Intelligence Augmentation)? Shen Xiangyang pointed out that intelligence enhancement (IA) refers to expanding human capabilities through technology to help people complete tasks more efficiently, rather than replacing humans.
In Shen Xiangyang’s view, one of the most difficult things in the big model is open source data. Therefore, the open source community needs to contribute more data. In the new paradigm, everyone can make greater progress together.
In January 2025, DeepSeek released the open source reasoning model R1. In addition to its powerful reasoning capabilities, it also has cost-effective features. Its recurring training cost is only 4% of o1, 0.2% of Grok-3, and the reasoning cost is only 18% of Grok-3 and 3.7% of o1-mini.Although the DeepSeek team did not attend the GDC opening ceremony, DeepSeek has become the key word in the audience.
“DeepSeek’s inclusive intelligence with limited computing power is a victory for open source and openness. rdquo; Qi Yuan, president of the Shanghai Academy of Science and Intelligence and Haoqing Distinguished Professor at Fudan University, said that open source will effectively accelerate the penetration of new technologies and lead to the large-scale application of AI technology.
According to Jevons ‘paradox, such high efficiency and low consumption will accelerate the speed and breadth of adoption of new technologies. Although the reasoning cost per token will be reduced, more people will use it after it is popularized, and the total amount will increase. Today, DeepSeek spends only millions of dollars in training costs and breaks ChatGPT in just 7 days to achieve a record of hundreds of millions of users in 2 months, making this Jevons paradox obvious confirmation and commercial success.
Qi Yuan said that this means that the reasoning cost of each large model token is getting lower and lower today, but as more people use it, the overall token usage may be higher, and everyone is using open source technology in large quantities at the bottom. On the other hand, people may also overlook trust. Open source itself can actually strengthen everyone’s trust in open source, and low cost will bring a very big and underlying commercial success.
It is reported that Qi Yuan’s team has developed a number of technical products such as Suiren Material Large Horizontal Model and Fuxi Meteorological Large Model. At the same time, Infinite Guangnian Company founded by Qi Yuan is committed to developing unique trustworthy large models and tool chain technologies, focusing on China’s horizontal and vertical product system and global development strategy: Yiheng Guangyu Qizhi is a one-stop intelligent computing service platform for training and promotion, and is China’s leading intelligent computing platform specifically oriented to science and intelligence; First,”Guangyu Golden Sail” precise reasoning large model and Agent application is an intelligent assistant specially created for financial professionals. Guangyu Golden Sail has been applied in leading financial institutions such as China Merchants Securities and HSBC Bank.
Qi Yuan pointed out that we are actually underestimating the future development brought about by the big model. Maybe new architectures will emerge after Transformer. Let’s talk about sparse architecture. A series of new technologies are believed to further promote the development of artificial intelligence. ldquo; The trinity of soil, moisture and sunshine will accelerate the development of science and intelligence. If we have an open ecology, it can have a very good promotion effect on universities, leading enterprises, and emerging enterprises.
“We talk about the Hacker spirit itself. The Hacker spirit itself is an open source and a victory for open source and openness. Scientific intelligence requires more diverse and open cooperation. The second value is amplified. Tokens are cheaper. In fact, more use is due to the rapid penetration of underlying technology, and the value of scientific intelligence will play a greater role.& rdquo; Qi Yuan said.
Jiang Daxin, founder and CEO of Step Star, said that since open source, the company’s products have received a lot of attention and praise, and creators around the world have used Step Star’s models to create massive videos. At the same time, more and more partners have joined Step Star’s open source ecosystem, including technology communities, creative communities, cloud manufacturers, chip manufacturers, etc. In March, Step Star will continue to open source graphics video products.
Shen Xiangyang emphasized that the biggest opportunity in the future lies in the relationship between man and machine, that is, human-computer interaction. He tends to define human-computer relationships as people-oriented artificial intelligence and calls attention to AI ethical issues.
“Looking back at the past four to fifty years, the development of the entire information society and the development of the entire computer industry, the most important thing is actually how people use computers. Which company grasps the entrance to the interaction between people and machines will become the greatest company in the world. In the PC era, products like Windows made Microsoft, search became obsolete in the Internet era, and TikTok was made in the mobile era. Now, in the AI era, DeepSeek is driving China’s AI leadership. rdquo; Shen Xiangyang said that in the process of developing artificial intelligence, we must keep in mind the people-oriented AI and pay attention to AI talents and the impact of AI on society.
Shangtang launches large model + large device applications to accelerate the construction of AGI platform
Under the DeepSeek craze, the development of domestic AI applications has become a key topic. At the opening ceremony, Yang Fan, co-founder and vice president of Shangtang Technology, mentioned a short story that triggered thinking.
“Speaking of landing application selection, this is actually a problem that bothers me. Shangtang has also done a lot of algorithms, computing power and applications, but now technology is developing too fast. I said, why don’t I ask about a big model that will be very popular in 2024? I said what key and major opportunities will be in 2025? It told me four things: expansion of application scenarios, breakthroughs in domestic computing power, improvement of data quality, and international competitive advantages. I asked DeepSeek, and it told me that the server was busy, please try again later.& rdquo;
Yang Fan believes that even though AI is very smart today and has more intelligence, it still helps us better learn and remember knowledge under the guidance of humans, and helps us better answer questions. How to better use AI tools and AI assistants in daily life and in the challenges of industrial development practice requires our continuous thinking, exploration and answers.
Today, relying on large model + large device capabilities to accelerate the construction of AGI goals is the main line of Shangtang Technology in the new decade, and implementation applications based on large models have also entered an accelerated period.
At the Shangtang sub-forum of the Global Developer Pioneer Conference on February 22, Shangtang Technology announced the new multi-agent large model application development framework LazyLLM, the Shangtang large device launch Vientiane platform, and the newCode Raccoon 2.0 and Office Raccoon 2.0 realize the transformation from Copilot to Agent.
Jia Anya, head of Shangtang Xiaoraccoon, said that the current breakthrough direction of AI can be roughly summarized into four aspects: how fast and economical it is, namely, more modal fusion, faster reasoning speed, better thinking ability and lower Training costs, which makes the implementation of large model applications enter an accelerated period.
Jia Anya believes that the main paradigms for implementing large model applications are Copilot and agent, which respectively represent different modes of collaboration between models and people in different application scenarios and different links. Shangtang Raccoon also happens to have two products: Code Raccoon and Office Raccoon, which are programming assistants for individuals and enterprises and office assistants for individuals and enterprises.
Among them, the new Code Raccoon 2.0 realizes multi-dimensional data fusion and multimodal reasoning capabilities based on the original code completion and question and answer interaction. It not only provides more efficient and intelligent code writing for C-end users and development tasks, but also provides enterprises with end-to-end software development solutions from requirements analysis, testing iteration to code asset management.In practical applications, Code Raccoon can help developers improve programming efficiency by more than 50%, and improve the efficiency of the entire enterprise R & D process by more than 30%.
The new Shangtang Office Little Raccoon 2.0 empowers more scenes by combining the powerful native fusion multimodal capabilities of the daily and new fusion model. Users can use multimodal understanding and interaction to better complete more general daily tasks, such as document processing, data analysis, etc., and can quickly realize vertical task analysis, research and final report generation.
GuShiio.comAGI learned that currently, there are more than 300 corporate users of Shangtang Code Raccoon and Office Raccoon, covering many industries such as Internet, finance, automobiles, manufacturing, etc., with more than 1 million direct individual users and more than 15 million final individual users. The number of daily tokens reaches more than 10 billion.
In terms of large equipment business,Shangtang Device released LazyLLM, a one-stop open source Agent application development framework for developers. This framework takes data as the core and supports continuous iteration of data during the application development process, thereby continuously improving data effectiveness. This framework can meet the specific needs of domestic developers for industries and specific domains, make up for the shortcomings of foreign tools, and ensure that software is independently controllable.
At the same time, Shangtang also released the Vientiane Agent Application Development Platform, which is a one-stop service platform for enterprise-level large-model application development and management. On the basis of the model management platform, it has additional functions such as visual orchestration, data management, knowledge base management, and tool management., Prompt management provides dozens of control canvas controls and resource controls, supports unlimited levels of canvas nesting, and supports multi-person collaborative working.
Relying on Shangtang’s device, the Vientiane Model Platform includes model management, fine-tuning, development, invocation, and application full-link services, providing a complete enterprise-level native development tool chain and complete technical service capabilities. The model development platform and the large model Agent application development platform can help the optimal implementation of the entire process of native AI applications. In addition, the Shangtang Large Device Vientiane platform has more than 5 self-developed large models of Shangtang ‘s daily new integrated large model system, as well as more than 500 high-quality preset large models.Including the latest DeepSeek series of models on shelves, the reasoning performance is 50% ahead of the industry, and the large model platform supports out-of-the-box use, which takes just 1 minute to get started.
At present, Shangtang Da Device Vientiane has extensively served customers in many industries such as central state-owned enterprises, operators, finance, Internet, automobiles, and pan-technology.With efficient model refinement and reasoning deployment, we can improve quality and efficiency in multiple business scenarios (improve quality and increase efficiency).
LazyLLM, the open source development framework for Shangtang Model, is essentially an open source AI application developer framework similar to LangChain+LLamaIndex.
Wang Zhihong, director of large equipment research and development at Shangtang Technology, told GuShiio.comAGI that the initial development of large models in 2023 will mainly focus on the model itself, including basic model training and model fine-tuning. However, as technology matures and market demand advances, more and more companies accelerate AI layout, demand has gradually changed from a single model to comprehensive application. In the past year, Shangtang has continuously explored key technologies for AI implementation, and combined its implementation experience to precipitate an AI application development framework LazyLLM to accelerate AI application development.
Wang Zhihong believes that 2025 is regarded as a key node for the implementation of AI applications, and the demand for AI models in many industries such as finance and medical care continues to grow. These successful cases not only verify the feasibility of the implementation of AI technology, but also provide digital transformation in all walks of life. Provide strong support. He pointed out that the number of customers is now increasing, and customers are gradually moving from training basic models or fine-tuning models at the beginning to implementing an application directly facing vertical landscape scenarios.
Talking about the impact of the DeepSeek craze,Wang Zhihong said that Shangtang immediately connected to the DeepSeek-R1-Distillation version of the reasoning service, and users directly access DeepSeek’s online services through LazyLLM. At the same time, the team also migrated DeepSeek’s related thinking chain distillation capabilities to LazyLLM, continued to increase LazyLLM’s application service experience, and gradually increased Proof-of-Concept (PoC) capabilities to 80%-90% in actual application scenarios.
“In essence, LazyLLM is not a platform that only does model training and reasoning. Instead, it integrates our various functional tools and new large models, and then uses my refined choreography to create a Pipeline (pipeline). The ability to provide application services to the outside world. rdquo; Wang Zhihong told GuShiio.comAGI.
Making good use of the DeepSeek model is part of Shangtang’s technical capabilities. At the same time, Shangtang will continue to develop new models every day.
According to the plan, in March this year, Shangtang Group will announce its latest annual report and progress based on the new strategic architecture of 1+X; at the Shangtang Technology Exchange Day held in April this year, it is expected to release the new Japan-Japan model 6.0 and a series of enterprise-level AI applications.
Currently, the AI boom is accelerating. With the continuous efforts of companies such as DeepSeek, the country is accelerating the development of the AI industry. By further improving the open source and open ecosystem, we will accelerate the promotion of large-scale applications in vertical fields and continue to improve the AI development level and core competitiveness.
Xiong Jijun, deputy minister of the Ministry of Industry and Information Technology, said that the Ministry of Industry and Information Technology will adhere to innovation-driven and application-driven, create a good development environment for developers, and provide strong support for the realization of new industrialization. The first is to make a good start in innovation and improve key technology innovation capabilities. The second is to improve the open source ecosystem and build an advanced open source service system. The third is to build an application test ground and accelerate the implementation and empowerment of technical products. The fourth is to build a strong reservoir of talents and stimulate the innovative vitality of global developers. The fifth is to integrate into the international stage and expand the space for global exchanges and cooperation.
