NextFin News — In June more than sixty million visits were recorded across seventeen Chinese desktop AI office agents. One product alone, Tencent’s WorkBuddy, accounted for nearly twenty-one million of them. ByteDance’s TRAE IDE followed with almost thirteen million. Alibaba’s earlier offering trailed at under eight million. Those numbers mark a clear change in how people use AI at work: the conversation has moved from answering questions to finishing multi-step tasks.
The same summer the three companies that dominate the traffic ranking each reorganized their agent efforts. Alibaba’s response arrived on August 3, when it opened public beta of QwenWork and released a new large model on the same day. The platform merges three earlier agents—one that controlled a user’s computer and files, one that ran longer tasks in the cloud, and one that worked inside enterprise messaging—into a single system. Chen Yusen, who took charge of DingTalk in mid-June, oversees the combined effort. The merger reached public testing roughly two months after his appointment.
QwenWork is available through a browser or a desktop program. Deeper links inside DingTalk are planned but not yet live. Users describe a goal in ordinary language; the system is designed to break the request into steps, use available tools, and return finished files rather than text replies. Specialized tool sets for product design or investment research can be added. Early connections allow it to summarize group chats, create documents and set calendar items inside DingTalk. Alibaba says later versions will reach company databases and internal workflows, though those links remain prospective.
The model released alongside the platform, Qwen3.8-Max, contains 2.4 trillion parameters in total but uses only a fraction for any single task. It can handle very long documents in one pass. Alibaba has shown demonstrations in which the model labels large numbers of legal clauses in about an hour and processes lengthy video into summaries in tens of minutes. Independent checks of those specific results have not been published. Pricing combines subscriptions with credits and offers separate plans for individuals and companies.
The traffic numbers and the consolidations point to the same underlying problem. Chat interfaces improved individual speed. They left largely unaddressed the harder question of organizational productivity—shared knowledge, access controls, multi-person processes and results that can be checked or reused. Each of the three large firms is now trying to move its agents closer to the software and data that companies already use every day. Tencent’s product leads in visits and benefits from easy login through existing messaging tools. ByteDance’s offering carries strength in coding-related work. Alibaba is betting on tighter links to its collaboration suite and on the ability to operate with company context rather than personal files alone.
Hu Yanping, a special professor at Shanghai University of Finance and Economics who studies intelligent technology, has described the digital workbench as a newly contested entry point. Integrating tools and data connections so agents can move across real workflows, he argues, is what turns a workbench into a company platform. The decisive factor is not model size alone but the ability to turn authorized company data and concrete processes into consistent results.
Research from Gartner has reached a parallel view. Differentiation among these systems, analysts there note, will depend less on access to the largest model and more on the capacity to convert company-authorized data, specific work scenarios and existing software connections into secure, reliable task completion. Model performance improves on short cycles. Understanding of how organizations actually work, and the ability to deliver inside environments with strict access rules, accumulates more slowly.
The market remains early. The June traffic figures show rapid adoption of desktop agents, yet they also show heavy concentration among a few products. Most of the measured visits still reflect individual use rather than deep embedding inside company systems. The technical ability to finish a spreadsheet or a presentation is advancing quickly. The institutional ability to handle permissions, audit trails, multi-user coordination and the idiosyncratic processes of different firms is advancing more slowly.
QwenWork is one attempt to close that gap. Comparable efforts are under way at the firms that currently lead in visits. The public tests now beginning will reveal whether any of these platforms can move beyond controlled demonstrations into the uneven routines of actual organizational work. Model updates can arrive in weeks. Turning agents into trusted infrastructure inside companies will take longer, and the outcome is far from settled.
Explore more exclusive insights at nextfin.ai.
