The 2026 Summer AI Technology Conference was recently held in Beijing. The conference focused on the release of a number of major technical achievements such as the new generation of large models and the Seedance series video models, disclosed the core operating data of the large model business, and demonstrated the AI implementation practices of benchmark customers in multiple industries.
From the core technical indicators ranking among the top in the world to the scale of industrial deployment achieving an order of magnitude leap, China’s AI industry is advancing in both directions of technological breakthroughs and scenario implementation, accelerating the transformation of artificial intelligence from technical verification to production-level large-scale application.
Large model capabilities have undergone qualitative changes, and core indicators have ranked among the first in the world
The newly released Beanbao Big Model 2.1 focuses on the three core production-level scenarios of coding, intelligence, and visual understanding to complete a comprehensive upgrade and officially breaks through the “production-level qualitative change point.”
In a number of international authoritative evaluations, the model has demonstrated first-class technical strength: in terms of coding and agent capabilities, the Terminal Bench 2.1 score is close to the level of the world’s top models, the SciCode evaluation scored 59.8 points, surpassing many international leading models, and the NL2Repo-Bench evaluation led the international similar products with 47.0 points, ranking among the first echelons in the world as a whole.
While maintaining its leading edge in multi-modal understanding capabilities, this large model has also achieved a significant reduction in usage costs: the professional version has a nearly 80% cost reduction compared to the world’s top similar products, and the lightweight version is only half the price of the professional version.
The upgrade of technical capabilities and cost optimization have opened up key bottlenecks for the large-scale popularization and application of large models in various industries. This means that domestic large models have officially entered the production-level application stage, and their capabilities are not reflected in demonstration effects, but can truly support core business processes: In a real customer scenario in the semiconductor industry, the model went through 18 hours of iteration to complete 1,300 lines of RTL code that can be directly put online; within technology companies, AI coding has been implemented on a large scale at the system level and has become a core link in the R&D process.
In the multi-modal field, the Seedance video model has also completed a leap in production-level capabilities.
Seedance 2.0 has achieved native 4K generation capabilities and supports 4K 10-bit high-bit depth native direct output.
The relevant person in charge revealed an iconic data: Before the launch of Seedance 2.0, the number of calls to the video generation tool on weekends was much higher than that on weekdays. It is essentially a tool with UGC entertainment attributes. After the launch of Seedance 2.0, the load and usage on weekdays exceeded that on weekends, confirming that it has entered the office production scene and completed the transition of productivity attributes.
At present, Seedance 2.0 has been implemented on a large scale in the fields of film and television creation, advertising, e-commerce marketing, etc. Film and television and short dramas are only one of the application scenarios of video generation technology, and will not be the core scenario in the long run.
At present, a large number of implementation practices have emerged in manufacturing, retail, education, autonomous driving, embodied intelligence and other fields: manufacturing companies use it to generate multi-language product descriptions and after-sales training videos; autonomous driving companies use it to synthesize extreme weather and edge road condition data to optimize algorithm robustness; embodied intelligence companies use it to synthesize training data to break the bottleneck of real data collection.
Video generation is an effective technical path to build world models. Relying on the powerful large model base, Seedance will penetrate into more core industrial scenarios in the future.
It is reported that the new Seedance 2.5 will be officially launched in early July, achieving world-leading breakthroughs in three dimensions: generation time, multi-material input, and editing capabilities.
Industrial penetration is accelerating and value implementation covers diverse tracks
Technology has crossed the qualitative change point and directly promoted the large-scale acceleration of industrial implementation.
Data disclosed at the conference showed that the average daily token calls of the Doubao model has exceeded 180 trillion, an increase of more than 10 times in the past year; in the public cloud MaaS market, its token share reached 49.5%; the number of members of the “Trillion Tokens Club” that deeply applies large models has exceeded 200, reflecting that enterprise-level AI applications are moving from pilot exploration to a new stage of in-depth production application.
At present, AI has been verified to create quantified value in multiple industries: coding is the earliest field to verify value globally, and the efficiency improvement effect of R&D has been widely recognized; horizontal scenarios such as knowledge base, knowledge Q&A, and data analysis have also achieved quantifiable value returns in a large number of enterprises.
In the vertical field, many global leading car companies in the automotive industry optimize the in-car cockpit experience through large-scale models to improve user satisfaction; leading companies in finance, education, smart home and other fields have achieved mature implementation; in the retail industry, some customers use AI tools to optimize store operations and consumer insights, and single store operation efficiency has been significantly improved.
In order to promote the deep integration of AI into enterprise production processes, relevant parties have built a product and service system covering all levels, and laid out a multi-agent product matrix from high-code to zero-code: professional developers can use professional development tools, internal developers of the company can build business intelligence through exclusive platforms, business personnel can use low-code tools to build applications, and functional positions can use out-of-the-box workbenches to cover efficiency improvement needs at different levels.
In response to the organizational adaptation problem in enterprise implementation, an industry solution team has also been established to co-create in depth with benchmark customers and tap the value of industry scenarios. At the same time, it has launched digital employee management capabilities and established an assessment, monitoring, and performance mechanism for intelligent agents to help enterprises adapt to the organizational management logic of the AI era.
In terms of overseas markets, MaaS access points have been deployed in South Asia, Southeast Asia, the Middle East, Europe and other places to support global customer access.
The relevant person in charge revealed that Seedance currently accounts for nearly half of its overseas calls. Many of the world’s top marketing groups and multinational companies have become customers. Model capabilities and cost-effectiveness are the core considerations for overseas customers’ selection.
Build a solid long-term foundation and anchor the development direction of industrial AI
Behind large-scale implementation is the dual support of underlying technical capabilities and compliance systems.
At the computing power level, we continue to promote self-research and optimization of the underlying infrastructure. Self-developed DPU, network switch and other technologies have been iterated for many years, improving overall computing efficiency through load offloading.
In response to the computing power requirements of different models, the relevant platforms have completed in-depth engineering optimization: the video model adopts a hybrid architecture based on Diffusion, which has lower requirements for high bandwidth and can efficiently adapt to multiple types of computing chips. There is no computing resource conflict with coding and agent models.
At the same time, the platform has completed the adaptation of multiple types of domestic computing power and can fully mobilize diverse computing power resources to support business implementation.
Safety and compliance are the core prerequisites for commercialization. The stronger the ability, the greater the responsibility. The construction of a safety system is the basis for model commercialization.
It is reported that Seedance 2.0 took several months from the completion of research and development to the official opening of the API. The core work is to improve the security strategy and IP copyright protection system, including face usage restrictions, copyright content identification, authorization mechanism construction, etc.
The IP business platform (preview version) launched at this conference uses a benign mechanism of formal authorization, electronic contracts, and profit sharing to open up the cooperation link between IP parties and users, allowing commercial creation to follow rules.
A complete security and copyright protection mechanism is not only an industry responsibility, but also builds a foundation of trust, which in turn will broaden the imagination of commercial applications.
When it comes to industry development, the relevant person in charge judged that China’s large model market is still in its early stages, but it has passed the qualitative change point of the production level, which is a crucial milestone. There may be room for growth of dozens or even hundreds of times in the future.
In the face of industry competition, the core barriers come from two aspects: one is the ability to continuously iterate models, and the other is the ability to implement in-depth industry services. The future foresight and strategic certainty of anchoring the direction of MaaS in advance also lay the foundation for the current market position.
Currently, the artificial intelligence industry is shifting from technology competition to a critical cycle of value realization.
The series of results released this time not only demonstrate the global competitiveness of China’s large-scale model technology in core indicators, but also confirm the broad space for the integration of AI technology and the real economy.
The relevant person in charge said: “The entire market is still in its early stages. Short-term wins and losses are not particularly important. What is more important is whether we can use better AI capabilities to serve more companies. As long as this is done properly, we will be able to achieve good results.”
