Region Focus | China

🌏 中文版

Region: China

China's office-agent war has entered its second act. In late July and early August, Tencent, Alibaba, ByteDance, and Baidu each independently ended their internal "horse racing" (letting multiple internal teams build competing products) and consolidated resources into single flagship offerings. This week, the three biggest players cashed in on that consolidation: ByteDance formally launched Doubao Work, Alibaba's QwenWork started expanding overseas, and Tencent's WorkBuddy demonstrated a more fundamental strategy -- not trying to have the strongest model, but making sure every model runs through its workbench. The same week's open-source release of GLM-5.3-Flash from Zhipu is exactly the kind of underlying condition that makes that strategy work.

Key Developments

ByteDance Launches Doubao Work, Folding in Feishu, TRAE, and Coze; Bloomberg Calls It a Direct Challenge to Tencent

ByteDance formally launched a new brand, Doubao Work, on August 25, positioned as an Agent product for office scenarios. Its most consequential design choice is account-level deep integration with Feishu: once a user signs in with a Feishu account, the Agent can directly inherit that user's existing chat history, documents, meeting notes, and schedule within their permission scope, without employees manually moving data over. Output files (PPT, Word, Excel) are saved back into Feishu cloud documents, where they can be shared and co-edited with colleagues via @-mentions. The teams behind ByteDance's coding agent TRAE and Agent-development platform Coze will be fully folded into the Doubao ecosystem. New users who download the desktop client or update the app get 30 days of free access, though ByteDance has not disclosed future pricing. Bloomberg framed the move as "ByteDance folding AI tools into its Doubao super-app brand to fight Tencent," while Caixin noted it continues a July reorganization that merged the Doubao and Feishu product teams and moved Feishu's enterprise sales into Volcano Engine. (Bloomberg, Caixin Global, Yicai Global, TechNode)

Alibaba Opens QwenWork International Public Beta, Deliberately Sidestepping Markets Dense with Western Incumbents

Alibaba announced on August 26 that QwenWork International has entered public beta, merging three of its Agent platforms -- QoderWork (software/web development), MuleRun (workflow orchestration), and Wukong (operating websites and local computers) -- into a single workspace, available on web and desktop simultaneously. The official announcement explicitly said the overseas push targets "fast-growing markets across Asia, the Middle East, and Latin America," rather than going straight after the Western enterprise market already dominated by Microsoft Copilot and Google Workspace. The interface initially ships in English and Simplified Chinese, with Traditional Chinese, Spanish, Portuguese, Japanese, and Korean to follow. It runs on a subscription-plus-credits model with Standard and Advanced model tiers; the Standard tier uses the newly released Qwen3.8 Flash model, which Alibaba says cuts per-task token consumption by 75% and roughly doubles generation speed. QwenWork only debuted in China as a public beta on August 3, and by August 17 was already ranked first among eight leading global workplace agents in an independent real-world evaluation by investment bank Jefferies. (Alizila, TechNode Global, The Reporter Asia, PANews)

Tencent's WorkBuddy Plays "Model-Agnostic": Collect Rent at the Platform Layer, Regardless of Which Model Wins

Tencent's WorkBuddy takes a fundamentally different approach from Doubao Work and QwenWork: instead of proving its own model is the strongest, it plugs DeepSeek, Zhipu's GLM, Moonshot's Kimi, MiniMax, and its own in-house Hy3 into a single workspace, letting users freely pick the underlying model, and captures a cut of every call through a points-plus-membership pricing structure. In other words, no matter which model wins, Tencent collects rent -- the only thing undisclosed is the take rate and whether the arrangement is currently profitable. Trade publication CIW News reports that since Hy3 rolled out free through WorkBuddy on August 5, its share among users who self-select a model has climbed past 60%. WorkBuddy's monthly visits grew from 8.85 million at its March launch to 20.97 million in June; combined with CodeBuddy and QClaw, the broader Tencent ecosystem reached 32.62 million visits in June -- more than half of the 60.62 million total the outlet tracks across the office-agent market. ⚠️ These visit and model-selection figures come from Analysys via 36Kr and from Tencent's own disclosures, and have not been independently audited. (CIW News)

Zhipu (Z.ai) Open-Sources GLM-5.3-Flash: MIT License, Runs on Domestic Chips, Priced at a Tenth of Its Predecessor

Zhipu's Z.ai released GLM-5.3-Flash on August 26, the first natively multimodal model in the GLM-5 series, with 320B total parameters and only 18B activated, supporting a 1-million-token context window. Architecturally, it is the first in the GLM series to combine sparse and linear attention (KDA linear attention plus NoPE sparse MLA), adding IndexPool to compress indexer key vectors -- Z.ai reports roughly 3x less attention compute and a 4.4x smaller KV cache. Pricing is $0.15 per million input tokens and $0.50 per million output tokens, about one-tenth of GLM-5.2, while scoring 84.3 on Terminal-Bench 2.1, closing in on Claude Opus 4.8. Weights are open-sourced under an MIT license on Hugging Face, and the model was trained and served entirely on domestic Chinese chips. Notably, the model spent the week before its official release running anonymously as "Ox Alpha" on OpenCode and OpenRouter's coding leaderboards, until forensic clues traced it back to Zhipu -- matching exactly the community signal this site tracked last week. This kind of "cheaper and stronger" open model keeps compressing prices at the model layer, which is precisely the underlying condition that makes Tencent WorkBuddy's aggregator strategy keep paying off. (Z.ai blog, MarkTechPost, Hugging Face)

Deep Analysis

The most noteworthy signal this week is that four Chinese platforms -- Tencent, Alibaba, ByteDance, and Baidu -- arrived at nearly the same strategic conclusion at nearly the same time: the durable moat isn't the model layer, it's the switching-cost and distribution layer. Applying moat theory (economic moats: switching costs, network effects, cost advantage, and intangible assets) to this week's developments reveals a consistent pattern:

Switching-cost moats are the real target of this consolidation wave. Doubao Work requires a Feishu login to unlock its full capability, and once bound, the Agent can read an entire organization's chat history, documents, and permission structure -- while everything it generates is written back into Feishu cloud documents. Replacing Doubao Work effectively means replacing the whole team's collaboration platform, and that migration cost is deliberately engineered to be high. Alibaba's three-in-one merger of QoderWork, MuleRun, and Wukong follows the same logic: get users to accumulate workflows and data in a single entry point before worrying about long-term retention.

Cost-advantage moats at the model layer are being destroyed by the platforms' own peers. GLM-5.3-Flash approaching last-generation flagship performance at one-tenth the price shows that cost advantages at the model layer now have a shelf life measured in weeks -- today's price edge can be matched by another open model next week. That's exactly why Tencent's WorkBuddy doesn't bet on having the strongest in-house model; instead it plugs DeepSeek, GLM, Kimi, MiniMax, and Hy3 into one workspace. Rather than compete on price in a layer that's being commoditized fast, it steps back a level and collects rent as the aggregator.

Network effects and intangible assets here come from existing social and collaboration infrastructure, not from model capability. Tencent leverages WeChat and WeCom's distribution reach so a user can issue a command from their phone and pick up finished output on their desktop. ByteDance leverages Feishu's existing identity and document graph. Alibaba plans to connect QwenWork with DingTalk. None of the three companies' intangible assets are model capability -- they're existing collaboration networks. That also explains Alibaba's choice of overseas markets: Asia, the Middle East, and Latin America haven't yet been locked in by Microsoft 365 or Google Workspace's existing network effects, making it easier to build switching costs there than in a head-on fight with Western incumbents.

Takeaways for Taiwan Founders

• Don't treat "using the newest, cheapest model" as a moat: this week's example of GLM-5.3-Flash approaching a previous flagship's performance at one-tenth the price shows a model-layer cost advantage can be matched within a week. If your differentiation mainly comes from model choice, assume that edge will be erased, and invest instead in higher-switching-cost assets -- data graphs, workflow integration, existing account systems.
• Enterprise Agent teams without an in-house model can borrow WorkBuddy's aggregator playbook: plug Claude, GPT, Gemini, and open models into one workspace, and replace "betting on a single strongest model" with "a routing layer plus user-data lock-in" so you get paid no matter which vendor's model wins.
• QwenWork's market-entry sequencing is worth studying: rather than directly challenging the Western market already saturated by Microsoft Copilot and Google Workspace, it picked markets whose collaboration-platform ecosystems aren't yet locked in (Asia, the Middle East, Latin America). For resource-constrained Taiwanese startups going global, that sequencing is more realistic than defaulting to a US-first strategy.

Takeaway

I used to think China's three tech giants were fighting an arms race in office agents -- whoever scores higher on benchmarks, whoever ships first. Looking at this week's developments, I now see that the "end of horse racing" consolidation in late July/early August (QClaw folded into WorkBuddy, three products merged into QwenWork, Feishu folded into Doubao) was really the same script playing out three times: all three companies arrived, almost simultaneously, at the conclusion that the office desktop is the entry point that's actually worth owning, while the model itself should be treated as an interchangeable part. Tencent WorkBuddy's "collect rent no matter which model wins" strategy states that logic most bluntly -- and open-source competitors like GLM-5.3-Flash, which keep driving model prices down, are effectively paving the road for that platform-layer play.

References

• Bloomberg — ByteDance Challenges Tencent With New Doubao Workplace AI Agent
• Caixin Global — ByteDance Consolidates AI Office Tools Around Doubao
• Yicai Global — ByteDance Launches Doubao Work as China's Tech Giants Pivot to Office AI
• TechNode — ByteDance launches Doubao Work with Feishu integration and 30-day free access
• Alizila — Alibaba Launches QwenWork International Edition
• TechNode Global — Alibaba opens QwenWork AI agent to global users
• The Reporter Asia — QwenWork International Edition Launches in Public Beta
• PANews — Alibaba's 'QwenWork' International Version Starts Public Beta
• CIW News — Tencent's WorkBuddy shows where China's AI money is moving
• Momenta Media — BAT's Big Push into AI Office Software
• Z.ai blog — GLM-5.3-Flash: Frontier Intelligence, Flash Cost
• MarkTechPost — Z.ai Releases GLM-5.3-Flash
• Hugging Face — zai-org/GLM-5.3-Flash