ByteDance reshuffles its AI business, Zhang Yiming refuses to take shortcuts
The AI industry is still far too fickle. Tech giants, who previously competed for user entry points, hardware parameters, and ranking positions to grab C-end user scale at the start of the year, have shifted their focus to delivering B-end productivity value, providing office assistants for enterprises and office workers.
Tencent's WorkBuddy is a typical example of the scenario-driven model strategy, leading the pack among various office agents all the way, and integrated the QClaw product center with WorkBuddy in July; after the "DingTalk Infiltration" incident at Alibaba, Chen Yusen, the former head of MuleRun, took over DingTalk. In August, a number of office AI products including QoderWork, MuleRun and Wukong were unified under the brand "Qianwen Office", ending the internal multi-product horse race mechanism; Baidu also recently launched the integration of office agents, merging the team and resources of its internal office product dodo into Baidu Dazi, and its office Agent business has switched to centralized operation.
ByteDance, which excels at "chat" and emotional value provision, has made the most drastic adjustment among the four giants, showing a pragmatic tendency as well.
At the end of July, the Feishu product team was merged into Doubao, while the sales, marketing and customer service teams were merged into Volcano Engine to build advantages in computing power and data. According to The Information, at the Seed all-hands meeting, Zhang Yiming, founder of ByteDance, also shared his views on model distillation, making it clear that closed-source models and open-weight models cannot be distilled, "Seed can accept temporary backwardness, but cannot accept eliminating this backwardness by distilling competitors' models."
These tech giants that cannot find commercial outlets among ordinary consumers and existing products have abandoned the internet traffic mindset and collectively returned to business common sense. Who has truly crossed the qualitative change threshold of production?
Breaking Out of the Dialog Box
Contrary to the rule of traditional internet products that marginal cost decreases as the number of users increases, the growth of large model users and call volume will directly drive computing power costs to rise exponentially, leading to a mismatch between user scale and monetization efficiency, and falling into commercialization dilemmas.
The "lobster craze" earlier this year saw AI evolve from passive response to active execution, attracting a large number of high-net-worth office workers to use related products, and also making the industry see the possibility of "AI breaking out of the chat dialog box".
Office software has entered the era of intelligent collaboration, and has begun to undertake complete workflows.
Anthropic launched the Claude for Microsoft 365 plug-in, embedding AI assistants into office scenarios; Microsoft added new enterprise-level AI Agent and GPU cloud PCs to Windows 365, and released the Copilot Cowork agent to bind AI capabilities to enterprise office and development workflows; Google upgraded Gemini Notebook and integrated it into the Gemini App, Search and Workspace office ecosystem.
These overseas cases also prove that the B-end is the main profit-generating battlefield.
Although OpenAI gained C-end popularity with ChatGPT, its operating loss exceeded 20 billion U.S. dollars in 2025; while Anthropic, which focuses on government and enterprise clients, has seen its B-end revenue surpass other segments, proving that the enterprise-level market is the carrier of large model commercial value.
The domestic market is also following this trend closely, as large and medium-sized tech companies have shifted from pursuing general-purpose models and focusing on chatbots to betting heavily on office AI.
According to the report "China Office Agent Platform Market Insight for Q2 2026" from Analysys, in June 2026, the total monthly visits of 17 mainstream desktop AI-native office agents included in the statistics have exceeded 60 million times, three times the figure in March.
A report from Kezhi Consulting shows that the scale of China's enterprise-level AI Agent market reached about 21.2 billion yuan in 2025, is expected to increase to 44.9 billion yuan in 2026, and is projected to exceed 332 billion yuan by 2029.
In the second half of 2026, Tencent, Alibaba and Baidu will collectively release integrated AI office-related products. The enterprise-level AI market has moved from the cultivation period to the rapid implementation period, with customer demands gradually maturing. In this context, ByteDance's thorough reorganization this time is not that abrupt.
Specifically, the Feishu product team is under Doubao; the marketing, sales and customer service (GTM) system is integrated with Volcano Engine to form a brand new ToB organization called "Creativity Service Platform". Zhao Qi is fully responsible for the product side, Tan Dai is in charge of the commercialization side, and Wu Yonghui coordinates the basic model R&D.
The trend of Feishu moving down to office scenarios and Doubao moving up to enterprise services emerged as early as the first half of the year.
In the previous round of competition for office applications, DingTalk and WeCom took the lead. Feishu, a latecomer born for efficient collaboration, delivered solutions to enterprises with the brand perception that "advanced enterprises use Feishu", and its ARR exceeded 200 million U.S. dollars in 2023. Without the wave of large models, Feishu would most likely still be deepening and enriching its business scenarios.
But the AI era has overturned the old logic. More than 90% of Feishu's new customers are purchasing Feishu AI products at the same time. The weight of instant messaging in office scenarios is gradually decreasing, as Agents can directly break down tasks, generate plans and perform initial executions. The needs of these advanced enterprises have changed accordingly, so Feishu naturally needs to re-prove its value.
On the other side, Doubao has over 300 million monthly active users and over 200 million daily active users, making it the largest AI application in China in terms of user scale, leading far ahead in the C-end market. As of June 2026, the average daily Token call volume of the Doubao large model has exceeded 180 trillion, leading to a sharp surge in computing power costs. Not satisfied with lightweight casual chatting and basic copy generation, Doubao has begun to explore commercialization, launching a professional version subscription service in June to enter office scenarios in a tiered manner.
Feishu becoming a division of Doubao dedicated to office scenarios also avoids internal resource waste, which is also reflected in the integration of the ToB GTM organization.
Previously, Xie Xin from Feishu was in charge of products, and also had independent sales, marketing and customer service teams. After the integration, his reporting level has dropped significantly, reporting to Zhao Qi instead of CEO Liang Rubo, which is in line with ByteDance's consistent talent strategy of prioritizing business performance.
In addition, according to data from Tianyancha and third-party research institutions, Volcano Engine's share in China's public cloud MaaS sector is close to 50%, but its accumulation in the IaaS sector is still far weaker than that of Alibaba and Tencent.
All commercialization efforts are centrally managed by Volcano Engine, with all B-end products sharing one sales team, shortening the decision-making chain internally and streamlining the delivery process externally. If the two sides can run in smoothly, the momentum brought by the organizational adjustment will be converted into corresponding market share, making up for the shortcomings in the ecosystem.
Rejecting the Distillation Shortcut
ByteDance rose to prominence relying on the dividends of traffic and algorithms in the past. The outside world's impression of ByteDance mostly stays in pan-entertainment scenarios such as Douyin, Toutiao, Hongguo and Doubao, and its hardcore technical strength is rarely the main topic of industry discussions.
With the in-depth reconstruction of this round of AI strategy and organizational structure, ByteDance's technical route choice has been pushed into the spotlight. Zhang Yiming's clear refutation of the "taking shortcuts through distillation" approach for large models has sparked extensive heated discussions in the industry. According to The Information, Zhang Yiming stated at the Seed all-hands meeting: "We are willing to sacrifice short-term gains for long-term goals."
CEO Liang Rubo also stated recently, "Some external reports said that we insist on self-developed models only because of external pressure, which is pure speculation. The real reason is that we hope the team can delay gratification and lay a solid technical foundation, which is critical for the long-term realization of AGI."
It has become a common practice in the AI industry for small models to reduce training time and costs through distillation, that is, learning and imitating large models, but distilling Claude can only get infinitely close to its performance instead of surpassing this ceiling. It is reported that Seed has discussed the feasibility of the distillation route many times internally. Now that the two senior executives have set the tone for ByteDance's independent R&D, different voices have emerged from the outside world.
Some commentators believe that ByteDance is worried about the recurrence of the TikTok censorship incident and leaving others with a handle, claiming that it is difficult to define the red line between reference and innovation for models; another voice attributes Doubao's lag behind other companies to its refusal to distill foreign large models.
This is essentially the contradiction between rapid iterative evolution and underlying capability breakthroughs in the AI industry.
Xu Xin, the renowned venture capitalist, revealed in a podcast that Douyin is a huge business, but Zhang Yiming now spends 50% of his time on Seed.
Seed is the technical source and strategic trump card of ByteDance's AI system. The Seed 2.0 large language model did not get very positive feedback, and it is not in the leading position in Coding, Agent and office scenarios. Liang Rubo proposed at the All Hands meeting that Douyin and Doubao should be taken as the two major "rough main lines", while admitting the gap between ByteDance's large language models and overseas leading players.
The Seedance series of video generation models is an exception, widely recognized as the world's highest-performance video model, maintaining the SOTA (State Of The Art) level. Volcano Engine set its overall 2026 revenue target at over 40 billion yuan, with its revenue structure highly dependent on video models, which also proves that ByteDance is not lacking in technical strength, and strong model capabilities can indeed be converted into commercial moats.
ByteDance, which lags behind in language models, also wants to take a gamble.
According to LatePost, ByteDance is discussing training a model with a parameter scale of over 5 trillion, led by Xiang Liang, head of Seed Foundation, and supported by Shen Ke, head of LLM pre-training data.
This model will exceed Alibaba's Qwen 3.8-Max (2.4 trillion parameters) and Moonshot AI's K3 (2.8 trillion parameters). As for whether this gamble will succeed, the final conclusion can only be drawn after seeing its actual implementation results.
The Last Mile
Shifting from "selling headcount" to "selling Tokens", tech giants have entered an all-dimensional competition covering model capabilities, data assets, tool ecosystems and delivery capabilities, until they truly solve the "last mile" problem of enterprise implementation.
Tencent's WorkBuddy has taken the lead with its product logic, with monthly visits reaching 20.97 million times in June, exceeding the sum of ByteDance's TRAE IDE (domestic version) and Alibaba's QoderWork, and its online and offline promotion efforts are still being strengthened. Alibaba's Qianwen Office and Kingsoft WPS AI are also trying to embed themselves into workflows to help users solve problems.
All parties are integrating their ToB businesses and standing at similar starting points. The boom of AI office has just begun, and it is not only ByteDance that has made changes.
References
[1] "ByteDance's AI Strategy Has Changed", Huang Qingchun Channel
[2] "ByteDiscusses Training a Model With Over 5 Trillion Parameters, Zhang Yiming: No Distillation, Don't Be Distracted by Short-term Hotspots Like Coding", LatePost
[3] "The Integration of Feishu and Doubao: A Goal Reshaping of ByteDance", GeekPark
This article is from the WeChat official account "Benyuan Finance", written by Li Youshan, and authorized for release by 36Kr.