ModelBest, China's largest unicorn in the field of edge large models, filed documents for consultation on an Initial Public Offering (IPO) with the Shanghai Securities Regulatory Bureau on August 12. CITIC Securities acted as the consulting institution.
The company completed a funding round on July 15, resulting in cumulative financing exceeding 5 billion yuan in the first half of 2026, and a valuation surpassing 20 billion yuan, making it the most valuable public company in edge AI in China.
ModelBest was founded in Beijing in August 2022 and incubated at Tsinghua University's Natural Language Processing Laboratory. Co-founder, Chairman, and CEO Li Dahai previously worked as a partner and CTO at Zhihu, where he was involved in product, technology, and commercialization for many years. Co-founder and Chief Scientist Liu Zhiyuan is a professor of computer science at Tsinghua University and a long-term researcher in natural language processing, pre-training, and large models.
The company's core focus is defined by the combination of an operator with deep experience in consumer internet scaling and an academic with extensive pre-training experience. The company's main business is the research and commercialization of edge large models based on the MiniCPM model family. These edge models run directly on devices—phones, cars, PCs, robots, and other local hardware—continuing to function under weak network connectivity or complete isolation where cloud solutions fail.
In 2024, the ModelBest team, together with researchers from Tsinghua University, proposed the Edge Large Model Act, which was published as a lead article in Nature Machine Intelligence in November 2025. The model series has accumulated over 38 million downloads on GitHub and Hugging Face platforms.
The fastest progress in commercialization is observed in the automotive industry. ModelBest's edge models have entered mass production in vehicles from Changan, SAIC, and Geely, and several hundred thousand vehicles are expected to be equipped with these models in 2026. Other deployment markets include smartphones, PCs, smart home systems, unmanned aerial vehicles, civil aviation, and the legal sector.
The combination of a robust fundamental research pipeline, a clean open-source distribution track record, and a clear reference to mass production in the automotive sector gives ModelBest a relatively unique profile for a Chinese AI IPO candidate.
The filing represents a registration at the consultation stage, not an issuance prospectus, and the path from consultation to official listing on the A-share market usually takes many months. The charter capital is 1.55 million yuan, and the largest shareholder, Beijing Qingyu Qihang Science and Technology Center, holds 16.45 percent. There is no single controlling shareholder. Current investors include national-level funds, central state-owned enterprises, automakers, and well-known financial investment companies, reflecting both the strategic importance of edge AI and the company's successful integration into real production environments.
The monthly active user count for Tencent WorkBuddy reached 20.97 million visits in June, surpassing the combined portfolio of all Alibaba artificial intelligence products. This situation put pressure on DingTalk's new CEO, Chen Yusen, who decided to launch Qianwen Office within 46 days of consolidating three existing products.
Ma Nan personally praised WorkBuddy during Tencent's quarterly report, calling it the most widely used AI efficiency agent in China. Furthermore, President Martin Lau mentioned WorkBuddy three times in the same call, while Yuanbao was mentioned only once, signaling a strategic priority for Tencent.
WorkBuddy's success is attributed to its integration with WeChat, including QR code login, mini-program integration, and setup without specialized knowledge. Non-technical employees use this tool for creating reports, presentations, coding, and participating in meetings. Tencent captured the ecosystem niche while Alibaba focused on internal competitive races, and ByteDance positioned Kouzi as a developer tool.
However, WorkBuddy faces structural limitations when deeply penetrating the corporate sector. WeCom functions more as a customer communication tool rather than an enterprise workflow operating system. Core business data, approval processes, and organizational rights are handled in DingTalk or Feishu, not WeCom. Qianwen Office, conversely, can directly interact with ERP interfaces, approval mechanisms, and permission systems within the closed DingTalk ecosystem. This allows for tasks such as financial reconciliation requiring ERP inventory data, budget comparison, multi-level approval, and generation of payment orders, all within a single seamless process.
WorkBuddy is forced to extract data from external ERP systems, process it, and then write it back, creating an additional step, delay, and security risk for regulated industries. Meanwhile, DingTalk serves over 26 million enterprise organizations with deep workflow integration, despite the paid conversion rate remaining below 1%. The competitive dynamic is shifting from consumer feature competition to the depth of integration into corporate workflows. WorkBuddy's strategy, based on expanding the skills ecosystem and gradually introducing enterprise functions through partnerships with ZTE, Meituan, and Ping An Bank, represents a long-term peripheral approach. In contrast, Qianwen Office's strategy, focused on deep integration with DingTalk, aims for the position of an enterprise operating system, where AI becomes an integral part of every workflow—from expense reimbursement to hiring and sales analysis.
The battle line runs between the advantage of WorkBuddy's consumer ecosystem and the depth of DingTalk's corporate integration. The ultimate winner will depend on which approach converts more effectively into paid corporate subscriptions. Tencent faces the challenge of gaining enterprise trust without an enterprise operating system infrastructure, whereas DingTalk struggles with the task of converting its massive base of free users into paying customers. Both approaches have yet to prove their effectiveness at scale for office AI agents, making the desktop office AI market the most significant strategic battleground in Chinese consumer-enterprise technology.
