Tencent’s WorkBuddy shows where China’s AI money is moving
Three giants consolidated competing desktop agent product lines in 14 days. The platform layer, not the model layer, is collecting the rent.
Tencent, Alibaba, and ByteDance consolidated their desktop agent product lines in 14 days in late July 2026.
WorkBuddy routes to DeepSeek, GLM, Kimi, and MiniMax, and earns points revenue on every call.
Office work is China’s largest token sink by disclosed revenue, powering ByteDance’s USD 4B model run-rate.
Hy3 (a.k.a. Hunyuan) went free worldwide on 5 August 2026, well ahead of any measurable overseas demand.
A desktop AI agent is not a chatbot. A user issues a plain-language instruction. The software plans a multi-step workflow, reads local files, calls tools, and returns a finished document, spreadsheet, or slide deck.
The category sits between two familiar products. Coding tools like Cursor serve developers while chat interfaces like ChatGPT answer questions. Desktop agents target the knowledge worker who does not need another conversation and wants a finished report or spreadsheet on the disk.
The window opened in early 2026. Alibaba’s QoderWork shipped in January. Tencent’s WorkBuddy followed in March. ByteDance later renamed TRAE SOLO to TRAE Work in July, pivoting a coding tool into the office race.
Model quality had crossed a threshold where multi-hour autonomous tasks became reliable, not merely demonstrable. StepStar chairman Yin Qi, quoted in Chinese by 36Kr, described the shift as models moving from seconds of continuous task performance to tens of hours of independent work.
China’s AI agent market reached RMB 80.4B in 2025, a 123.2% YoY increase, per iiMedia Research. Analysts project RMB 696.8B by 2030. Five months into the desktop agent category, the horse race is over.
14 days that ended the race
Between 20 July and 3 August 2026, three companies retired competing products in a coordinated retreat.
Tencent moved first. On 20 July, its QClaw agent business folded into the Cloud Product Division that houses WorkBuddy, according to 36Kr. QClaw continues to operate under its own brand for now. The team and budget consolidation is complete.
Alibaba consolidated three products into one. QoderWork, Wukong, and MuleRun merged into Qianwen Office, which opened public beta on 3 August, according to 36Kr. Chen Yusen, the new DingTalk CEO, runs the combined business unit. Group CEO Wu Yongming reportedly attends key meetings personally.
ByteDance took the most radical step. On 30 July, an internal email from CEO Liang Rubo merged the Feishu product team into Doubao. Feishu’s sales organization moved to Volcano Engine. A standalone product line effectively disappeared inside a chat brand.
The switching cost dynamics explain the urgency. Installing a desktop agent means granting folder access, binding accounts, and embedding the tool in daily workflows. Casual replacement becomes impractical. The product that occupies the primary window at scale captures the addressable market for years.
Coordinated retreat signals a closing window. No company times a full reorganization for July unless the market rewards speed. All three concluded that parallel product bets no longer paid, and that the desktop entry point rewards first movers.
What the visit data does and does not prove
Analysys published its “2026 Q2 China Office Agent Platform Market Insight Report” on 20 July, according to 36Kr. The methodology deserves attention before the numbers do.
The count measures PC-side web visit volume across 17 mainstream desktop agents. Daily active users, paying enterprises, and revenue are not counted. Mobile usage is excluded.
Total June 2026 visits reached 60.62M. WorkBuddy alone drew 20.97M, up from 8.85M at its March launch. The Tencent ecosystem, including WorkBuddy, CodeBuddy, and QClaw, totaled 32.62M, more than half the tracked market.
ByteDance products reached 14.41M, driven by TRAE IDE. Alibaba’s QoderWork and Wukong totaled 9.19M. The three ecosystems combined captured 56.22M of the 60.62M visits counted.
One caveat matters before drawing conclusions. TRAE IDE is a coding agent, not an office agent. Its inclusion inflates ByteDance’s ranking, and an Agent developer quoted by 36Kr called the placement misplaced on those grounds.
Even after discounting TRAE IDE, concentration remains real. The category is not a startup market.
Desktop visit growth still means something specific. A user opening a PC-side agent is committing to a workflow, not idly browsing. Switching costs run higher than a mobile app tap, which is why every player is racing for the entry point rather than the model.
How Tencent monetises a model it did not need
Tencent does not build the best model in China. Alibaba’s Tongyi Qianwen leads most benchmarks, Moonshot’s Kimi holds a long-context edge, and DeepSeek redefined training efficiency.
WorkBuddy sidesteps the comparison by operating as a model aggregator. Users route requests through Hy, DeepSeek, GLM, Kimi, or MiniMax, according to Tencent Cloud. A points-plus-membership pricing structure captures revenue on every call, regardless of which model runs underneath.
The economics matter more than the technology. When Kimi wins a task, Tencent earns platform commission. When Hy3 wins, Tencent earns platform commission plus first-party model revenue. The house holds no losing position.
One number remains undisclosed. Tencent has not published the take rate on third-party model calls, whether those calls are subsidised, or whether aggregation is currently profitable. The mechanics point clearly at platform-layer capture. The unit economics are asserted, not proven.
Hy3 is now taking share inside the platform. Among WorkBuddy users who select their own model, over 60% choose Hy3, according to Analysys via 36Kr. Tencent released Hy3 on 6 July 2026, then made it free worldwide through WorkBuddy on 5 August, per the Tencent newsroom. The model runs a Mixture-of-Experts architecture with 295B total parameters, 21B active, and a 256K context window, available under Apache 2.0.
Jiashi Consulting partner Li Yingtao, quoted by 36Kr, categorises Tencent’s position as a traffic-entry-point play, with WeChat and WeCom distribution driving low customer acquisition cost. Distribution is the moat. A WeChat trigger inside WorkBuddy lets a user start a task on a phone and pick up the finished output on a desktop.
Value capture happens at the desktop, not the model. Aggregators historically win when the underlying commodity has many suppliers and no clear technical moat. China’s model layer, with multiple credible open-weight contenders across DeepSeek, Qwen, Kimi, GLM, and Hy, matches that description. WorkBuddy is priced to collect rent on the choice, not to win it.
Why office work is the token prize
The consolidation was not defensive. The July disclosures merely confirmed to the market what the three participants already knew from their own token metering.
ByteDance disclosed in its 30 July internal email that its large model business run-rate exceeded USD 4B in annualized revenue, calculated on average July 2026 consumption, according to 36Kr. Doubao monthly active users reached 382M. Both figures come from company disclosure and lack independent verification.
