Summary

The defining move in 姚舜宇’s first 300 days after parachuting into Tencent’s Hunyuan was not a model launch but an organizational purge. During the first 2-3 months, when he was at the peak of the boss’s trust, the heads of pre-training, post-training, evaluation, and infra were “basically replaced” with people he recruited himself; the entire infra team came from ByteDance’s Seed infra, regarded as the industry’s strongest. 洪灏’s maxim is worth remembering: “The smartest reformers do the most radical things when they have the boss’s greatest trust.” Tencent’s boss may be patient for “2-3 years.”
Hunyuan 3, a roughly 290B MoE, met expectations rather than surprising anyone. It was internally positioned as a small model: “The core goal is to get the whole organization working together and restore team morale.” A preview in April was used to manage expectations, with the full release moving up one level. 姚舜宇’s own assessment is clear-eyed: Hunyuan 4 “may not truly enter the first tier” either; a genuinely influential model will come later. The catch-up narrative—“surpassing Qwen is only a matter of time”—is being sustained through release pacing and expectation management.
姚舜宇 chose Tencent based on the argument in his blog The Second Half: the next phase will be about “defining the problems truly worth solving,” which depends on a large company’s context. In his view, Tencent and Meta have the most context; ByteDance’s 800M DAU is offset by “homogeneous user profiles,” while an ecosystem built around recommendation and distribution was passed over.
The context thesis has collided with reality: WeChat still has not given Hunyuan access to its data. 张小龙 is “extremely concerned” about WeChat’s private user data and would rather build a controllable 258B model in-house, making it likely that the 2 foundation models will coexist for the long term. With Chinese companies operating at least an order of magnitude behind overseas leaders in compute, whether Tencent’s “federal system” can pull together is Martin’s central question. Internally, the metaphor is Black Hawk Down: Delta Force executes the capture perfectly, the Rangers responsible for coordination drop the ball, and the mission ends in disaster.
The highest-increment deployment initiative is a company-wide reinforcement-learning platform. Hunyuan is assigning post-training teams of 2-3 people to each business, with Work Body and Yuanbao as seed users; much of the data used to train Hunyuan will come from Work Body. Linking the model layer with the application layer is the second battle after the organizational overhaul.
The C paradox is the episode’s sharpest investment debate. Everyone assumes ByteDance’s C team will stay in the game because it has the best resources, talent, money, and infra, yet it has not produced the kind of reputation attached to Kimi, DeepSeek, or Qwen; in coding, it “may not even be at the top of the second tier.” 曼祺’s explanation is that this confidence is merely “a memory of short-term historical momentum,” an essentially emotional source. The observable signal is that people leaving ByteDance raise money more easily.
Coding commercialization is running into cold water. Enterprises buy by seat, but actual seat utilization is low; “whether you look at selling by seat or at C-end payments, neither is particularly optimistic.” The growth slope from programmers to white-collar users remains “very low.” 洪灏 worries coding will become SaaS—slow growth, weak revenue, and no winner-take-all dynamic—and asks whether “the imagination space in China is simply quite limited.” So far, commercialization has only been validated in coding and video generation; meanwhile, “since 2023, every autumn people have said the bubble was about to burst,” and it has happened 3 times.

Deep dive

1. After DeepSeek’s Shock, Tencent Bets on Youth

DeepSeek’s breakout in early 2025 convinced Tencent executives that “an organization with relatively little experience and full of young people could actually build such an impressive model.” Their own team had aged, with many people coming from traditional search, advertising, and recommendation backgrounds—the same path ByteDance followed early on with 朱文佳 and 乔木. The strategy was set: find young people.
Internally, 2 paths were considered: gradually replace the old team, or build a separate young team to race against Hunyuan. Tencent ultimately kept Hunyuan’s existing structure and replaced people step by step. Martin (刘炽平) went to Silicon Valley to meet researchers at Meta and Google; research interviews, previously conducted “up to the VP level,” began to include Martin himself.
洪灏’s comparison: ByteDance moved earlier, with 张一鸣 meeting researchers intensively more than a year ahead of Tencent. The “young people trend” was already visible in 2023—曼祺 recalled reports that quoted investors refusing to meet founders over 35—but it spread into an industry consensus only after DeepSeek.

2. How 姚顺宇 Landed at Tencent: Introduced at a 2024 Conference, Deal Closed in Fall 2025

Tencent’s senior-hiring team met 姚顺宇 at a major conference in 2024 and introduced him to senior management, but nothing came of it for a long time. In 2025, for various reasons, he wanted to return to China and had “contacts and conversations with several major tech companies”; he decided to join Tencent in fall 2025 as head of Hunyuan. 洪灏 acknowledged that he did not know whether 姚顺宇 had considered starting a company.
He reports directly to 刘炽平 and also to 卢山, president of TEG. 卢山, however, primarily oversees multimodal work and has limited involvement in the Foundation Model.

3. Why Tencent Rather Than ByteDance: The Second Half and the Context Debate

姚舜宇’s April 2025 blog, The Second Half, argued that “the next phase will no longer be a competition in training techniques, but in defining the problems truly worth solving.” That requires the context accumulated by large companies. In his view, Tencent and Meta have the most: social data; workplace scenarios across WeCom, Tencent Docs, and Tencent Meeting; QQ Browser and WeChat Search; Tencent Video and QQ Music; and games—“every vertical has an application that has reached the top tier.”
ByteDance was passed over because Douyin has 800M DAU, Toutiao around 100M, and Hongguo and Fanqie around 100M each, but the user profiles are homogeneous and the ecosystem revolves around recommendation and distribution. The diversity of its data is “far inferior to Tencent and Meta.” 曼祺 added that Feishu has unusually valuable data accumulation, but its volume is far below DingTalk and WeCom.
曼祺’s reading of the blog: the first half conquered intellectual games with clean evaluations—Go and the IMO—not real production tasks. The second half requires finding the right tasks for AI, with the technical prerequisite that RL has already demonstrated generalization.

4. Hunyuan’s Prehistory: Born from a Model Race, Starting with 2,000 Borrowed Cards

Hunyuan emerged from a race among multiple teams inside TEG. AI Lab, founded in 2016 and long focused on NLP, was the group with the deepest large-model expertise but failed to deliver. A team from the data platform division won instead and became the base for Hunyuan 1 and 2; departments were then combined into a “joint force.” Most members came from the previous generation of search, advertising, and recommendation engineering.
Cost-cutting and GPU order reductions in 2021-22 left Hunyuan starting with just 2,000 cards, borrowed from the advertising division. Advertising could afford GPUs because it was a revenue-generating business and was still building ad-recommendation models on the BERT playbook. “This was a somewhat crippled organization.”

5. Researchers Arrive at the End of 2024: What Was Missing Was a Product-Manager-Like Sense of Direction

Tencent brought in 谭旭 from Microsoft via Kimi, with expertise in speech and multimodality; 冯佳时 from ByteDance, also in multimodality; and 彭厚文 from Microsoft Research Asia, who has recently been rumored to have left. The diagnosis was that Hunyuan had AI engineers but no native AI research team—“like building a product without a product manager: you don’t know what to do next, and much of the work feels like busywork.”
曼祺 highlighted the timing gap: ByteDance only fully realized “this thing worked” when GPT-4 launched in April 2023, which was already late; “Tencent was even later than them.” It was consistently half a step slow from recognition to action.

6. The Purge: Every Key Role Replaced in 2-3 Months—“Don’t Touch Model Work; Do Whatever You Want”

姚舜宇 “knew that he had to replace every important role quickly while he had the boss’s trust”: the heads of pre-training, post-training, evaluation, and infra were essentially all replaced, with every hire made by him. The infra hires came from Seed, considered the industry’s strongest team, including 肖学峰, 张弛, and 黄启; 刘辉丹 joined pre-training.
Most of the old hands stayed at Tencent but left Hunyuan, were sidelined, and transferred out on their own. A Hunyuan contact described 姚舜宇’s approach: “You can keep sitting here. As long as you don’t touch model work, do whatever you want.” When 曼祺 asked whether that counted as gentle treatment, 洪灏 disagreed: “Replacing a whole wave of people within 2-3 months is still extraordinarily fast and decisive.” Some veterans, such as 徐谦, built trust with 姚舜宇 early through “excellent political sensitivity.”
In April and May, he also folded multimodal understanding into the large-language-model division, and his scope kept expanding. 洪灏’s off-the-record assessment was that there might not be a final plan: “This is a natural process that can happen—when the boss trusts you most, you can do a great many things.”

7. Tencent Opens Its Wallet: Hunyuan’s Appeal Is That It Has “Potholes Everywhere”

Hunyuan had previously paid below market. From 2025, Tencent decided to pay up for talent, with total compensation “in some respects higher than ByteDance and Alibaba” because of its relative disadvantage. The more structural pitch was that ByteDance and Alibaba already had mature systems, while Hunyuan was full of potholes: “You can come in and be a leader; you don’t have to go into C’s team and become another screw.”
The appeal to campus hires and young researchers was “very strong—sometimes even too strong for the team to hire everyone,” and 姚舜宇’s name itself helped recruit. The clearest exception to Tencent tradition: Hunyuan’s current architecture head is a PhD student. Historically, important business leaders at Tencent had to be deeply seasoned; a move like 韩尚佑 becoming Douyin’s head at 30 “would be very difficult at Tencent.”

8. Team Culture: Startup Energy Coexists with the Divide of “Only the Newcomers Are Smiling”

Multiple interviewees saw the outline of a startup team: fewer layers, transparent sharing of context, faster reactions, and a shared mission across businesses. But big-company problems remain. Hunyuan is far larger than Kimi or DeepSeek, so “there are definitely people coasting”; many arrived guarded, and names and responsibilities were opaque—“when you have an idea, figuring out whom to tell means making the rounds.”
The divide is between old and new: “the newcomers are the only ones smiling.” Veterans have seen their scope cut and their traditional ways of working stop working, which “must feel completely different.” 洪灏 confirmed that this is precisely the direction the company wants.

9. Split Verdict on 姚舜宇: Not Technically Sharp Enough, Remarkably Strong Organizational Instincts

The prevailing view among colleagues is that he is not technically driven and does not offer especially sharp views on technical direction. Unlike 梁文锋 and 杨志玲, who focus almost exclusively on technology, he delegates technical decisions and concentrates on the organization. He often says internally: “For large models, the organization is extremely important; many problems are organizational problems.” External researchers echo the view that he is “not as technically sharp as people imagine” and does not go as deep into details as 梁文锋.
Both interviewees said they were surprised that the 1997-born, now 29-year-old executive was “much more mature than expected.” “To care about the organization, how to prove yourself quickly inside it, and how to secure scope—without some senior person guiding you, a young person may not think about all that.” 曼祺 added that even with guidance, “you still need the instinct for it.”
Another repeatedly cited trait is his concern for the experience of people doing frontline execution. He regularly seeks feedback from the youngest researchers and interns and pushes rapid fixes—issues such as “the data platform is hard to use” receive immediate attention. 洪灏 sees this as “a form of organizational instinct: the source of information matters.”

10. Martin’s Mandate: Money, People, Compute, and Time—Reform Must Happen at Peak Trust

姚舜宇’s biggest sponsor at Tencent is Martin. Everything a reformer needs depends on trust: the money to hire the best people and replace the old guard, a time window in which the boss promises not to inspect benchmarks, and compute. Tencent’s compute allocation is essentially negotiated by TEG’s 卢山 and Davis (林锦华), who oversees the strategy and advertising lines; both ultimately report upward to Martin.
洪灏’s full framework is worth preserving: “A boss’s commitment is only a commitment. If he gives you 2 years of patience, you have 2 years of trust… The smartest reformers do the most radical things when they have the boss’s greatest trust.” Otherwise resistance only grows: “You keep proving that you’re right, trapped in self-justification, which is extremely painful for a reformer.” He estimates Tencent’s boss may be patient for “2-3 years.”
Communication is frequent enough. The executive office meets monthly; starting in 2025, Pony required Hunyuan-related directors and GMs to report weekly. 姚舜宇 is often in Hong Kong with management, so “it is entirely possible they communicate every day.”

11. Hunyuan 3: A Textbook Exercise in Expectation Management

Internal expectations for Hunyuan 3—roughly a 290B MoE, “a little over 290B”—were never high. It was a small model with limited problem-solving range; the point was “to get the whole organization working together and restore team morale.” The result met expectations rather than exceeded them, and the bosses at the executive office considered it acceptable.
The expectation management was deliberate. A Hunyuan 3 preview was released in the April 20s—“very clearly a version designed to manage expectations,” for both management and the team. The full release showed a relatively clear improvement over the preview, establishing the cadence. 姚舜宇 was also clear-eyed: Hunyuan 4 may not truly enter the first tier. The gap is narrowing, but “it may be further down the road before we produce a genuinely influential model.” Hunyuan 4 will add multimodality, expand text scale, and improve data; the entire infra stack was rebuilt from scratch in 3-4 months.
The pressure was already intense. When he arrived, the internal line from top to bottom was: “We know the problems, the people are here, everything is ready, and surpassing Qwen is only a matter of time.” Colleagues saw him as composed, methodical, and forceful—“strong internally and externally,” rather than bluffing. He dealt with the pressure by organizing basketball games with colleagues.

12. Two Camps After K3: Latecomers Can Catch Up vs. Who Is Still Building Small Models?

After Kimi K3 launched on Friday, July 17, and delivered an impressive showing, the industry’s view of Hunyuan’s pace split. One camp argued that large models “each have their moment for a few months,” and that latecomers can leapfrog with more advanced architectures, distillation from stronger models, and better data. 曼祺 joked: “Is Hunyuan going to distill?” The other camp asked: everyone is pushing toward the limits of larger models, and ByteDance is rumored to be training a 10T model—“what is the point of building this kind of small model?”
洪灏’s defense of small models was practical: new hires need to ship something within a reasonable period; “you can’t come in and have nothing to show after a year.” A completely new team also needs real-world practice to work through every link and every person.

13. Data Silos: WeChat Still Hasn’t Shared Its Data; Breaking the Ice Requires Proving Usefulness

The current picture: WeChat data “should be described as not yet shared” with Hunyuan. Tencent Video’s rights sit with China Literature and must be bought with real money. Tencent News has provided some data, but only for standalone projects—it “cannot possibly be used for pre-training”—and compliance remains an issue. “Data is each business unit’s most valuable asset.”
The proof point for breaking the silo is demonstrating usefulness first. After Tencent News integrated Yuanbao, “user activity rose very quickly”; interaction between the comments section and Yuanbao became common. Peacekeeper Elite’s collaboration with Hunyuan 3 on AI NPCs produced “very clear results.” Operationally, Hunyuan is assigning 2-3-person post-training teams to most businesses as dedicated interfaces, while the business units provide their own data.
洪灏’s framework: previously, every business was growing and needed no outside help. Now “every business has a very strong sense of AI anxiety,” so attitudes are much more open than before. But the message is still: “If you are just as bad, that means you have wasted my data.” Mutual trust will have to be built incrementally.

14. Product Portfolio Consolidated in CSIG: Yuanbao Pai’s 30,000 DAU Embarrassment

Nearly all of Tencent’s flagship AI applications now sit in CSIG: the workplace agent Work Body, coding product Code Body, Yuanbao as a chat box, knowledge base ima, and the App Store team’s desktop agent Marvis, which is deeply integrated with Microsoft and has broad PC permissions. QQ Browser, Sogou Search, and Tencent Docs were also moved from PCG into CSIG. The consolidation came from the boss wanting someone to “connect these products and fight this battle”; 汤道生 stepped forward. 曼祺 joked: “Did 汤道生 take a step forward, or did everyone else take a step back?” The answer: “Hard to say.”
The counterexample is 元宝派, a WeChat-Moments-like product inside Yuanbao: DAU is around 30,000, with only 1-2 people left maintaining it. “When a big company discovers it chose the wrong direction, it won’t directly admit the mistake and kill the product; it lets it slowly wither.” Cutting losses quickly is hard for large companies.
The product map is overwhelmingly focused on productivity, with entertainment and companionship embedded in existing apps: Peacekeeper Elite’s Oasis for UGC game creation, TME’s music generation, and Tencent Video’s Huo Long AI comics platform. On the model side, Hunyuan covers LLMs, video, image generation, and 3D; 姚舜宇 is responsible for dynamic understanding, while video generation, image generation, and 3D are outside his remit. IEG’s games division is building a separate world model “because it has the assets.”

15. Biggest Incremental Move: A Reinforcement-Learning Platform for the Entire Company

姚舜宇 is pushing a company-wide RL platform that every Tencent product can use for post-training, making the business-model loop more organic. 曼祺 considers this the most incremental part of the reporting. The plan is to push it again this year, but it has not yet been deployed at scale; Work Body and Yuanbao are the designated seed users because they are most closely tied to CSIG.
Work Body is led by Betty, formerly of Tencent Cloud; Yuanbao is led by 吴祖荣, who also heads Tencent Meeting. Both are regarded internally as people who can deliver. During the Lunar New Year chatbot battle, 姚舜宇 spoke with the Yuanbao team almost every week. Work Body is especially important: much of the data used to train Hunyuan will come from Work Body.

16. WeChat VLM: 张小龙’s Obsession with Control Means 2 Foundation Models May Coexist Long Term

Hunyuan and WeChat “have no direct relationship—basically no relationship.” WeChat’s agent Xiaowei is still in gray release and “probably supports only around 1M in scale,” routing by scenario among WeChat’s in-house model, DeepSeek, and Hunyuan. Most of the traffic from tagging Yuanbao in comments now goes through Hunyuan, and Yuanbao has not integrated DeepSeek V4—it still uses V3—“which shows that they still want to push Hunyuan.”
Martin has discussed the question of a common foundation model with 张小龙. His position is that he does not regard Hunyuan as a third party, unlike DeepSeek, Qwen, and Kimi, but “he is extremely concerned about WeChat’s own private user data and still prefers to build a model himself that he can control.” Internally, the teams are exploring federated learning and other ways to avoid data-privacy risks. 洪灏’s view: the 2 teams will very likely coexist for the long term.
The VLM team released a 258B MoE in January, slightly smaller than Hunyuan 3. It does not offer an external API and is used in Xiaowei, WeChat Search, and other native scenarios. On July 14, it updated its technical blog on “shallow-space compute scaling,” with the central problem framed as “extreme resource efficiency in an environment of 1.4B people.” The team is led by Harvey (周浩), a technical executive who has been at WeChat since its founding. Internal talent flows in both directions, but Hunyuan pays far more than the VLM team; WeChat’s compensation system is hard to change and has traditionally favored internal development. The team also approached talent such as 戴子航（音）from Grok this year, but 张小龙’s standards are high and cultural fit is difficult: “Hiring someone like that as the top executive is not easy.”

17. Xiaowei in Practice: Dancing Within WeChat’s Rules

张小龙 is now “very involved” in large models, and the team is “generally somewhat anxious.” WeChat is a 1.4B-person mass-market product whose rules are already fixed: “You can only dance within WeChat’s rules. How to break out of them and truly become AI-native is genuinely hard to say.” There is no clear internal view on the direction; “it really is being left to the users.” WeChat is an extremely user-oriented team.
洪灏’s hands-on experience: functions closely tied to chat are usable—explaining words, summarizing articles and group-chat records—but “my own need isn’t that large.” Using an agent to order food through a mini program is “too inefficient—by the time it finds what I want and confirms step by step, I have already ordered it manually 3 times.” Payment still requires human confirmation. Asked by 曼祺 what value such data has for the model, 洪灏 replied directly: “This kind of data may not have much real utility.”
The analogy is 小蓝包: “highly imagination-defying and highly native to WeChat.” But e-commerce is dirty, labor-intensive operational work. WeChat wants to achieve a lot with very little effort, and “that is where the gap appears.” With too few products, “it definitely cannot become mainstream e-commerce.”

18. Federalism Hits the Compute Wall: The Need to Pull Together

Tencent operates like a federal system: business groups have substantial autonomy, while the center retains influence but does not force uniformity. Games can build their own world model; Hunyuan also works on 3D and world models; “senior management will not impose mandatory requirements.” 曼祺 identified the new constraint: compute. Chinese companies are concentrating all available resources yet remain at least an order of magnitude behind overseas leaders, and the gap “may be 2 orders of magnitude” further down the stack. It is a serious problem.
洪灏 said Martin clearly believes Tencent needs to pull together, and that the conclusion predates AI and began with ByteDance’s rise. ByteDance’s core strength is “pulling all resources into one rope, connecting all data, and making everything circulate.” Tencent has reflected that its vaunted cross-department collaboration can become “extreme sharing of leadership”; facing a top-down, fast-moving rival, it needs unified command.
The internal metaphor is Black Hawk Down: Delta Force executes the capture perfectly, but the supporting Rangers get caught up in their own fight and the mission fails. The comparison is to the US operation against Maduro this year: despite using air and naval forces, it was driven entirely top-down by Delta Force and succeeded in one try. 曼祺 laughed: “Tencent seems to have studied Delta Force extensively.” The answer: “After all, it made Delta Force: Hawk Ops.”

19. A Highly Upvoted Claim That “Headquarters Can’t Get Big Things Done”—and the Real Problem of Edge Innovation

A highly upvoted comment under the article argued that WeChat came from Guangzhou and Honor of Kings from Chengdu, while headquarters’ microblogging, e-commerce, search, cloud, and Weishi all failed. Weishi was even led by 任宇昕 and had more resources than 姚舜宇. It concluded that 姚舜宇’s twin strategy of strong technology and top-down group control “violates Tencent’s genes” and is “destined to fail.” 洪灏 first corrected the facts: QQ, Tencent.com, Tencent News, QQ Music, and Peacekeeper Elite are all counterexamples, and Peacekeeper Elite was built in Shenzhen. “There are quite a few counterexamples; the claim does not really fit the facts.”
He did retain the underlying issue: the innovation dilemma of large companies. When 张小龙 was asked why WeChat was born in Guangzhou, he said: “Precisely because we were in Guangzhou, we didn’t need to know what people in Shenzhen were doing, and could work more peacefully according to what we believed was right; otherwise WeChat might never have happened.” One of the central ideas in Out of Control is that innovation occurs at the edges. Qwen was also lightly managed inside Alibaba early on; resources were only heavily concentrated after 2023. Hunyuan is based in Shenzhen, with some staff in Beijing.

20. The Reform-Cycle Thesis: The Next 300 Days Will Be Harder; 陆奇 Is the Cautionary Example

洪灏 sees Tencent AI through the lens of parachuted-in reform having a cycle. Today, 姚舜宇 is in his highest-trust period: personnel, organizational, and strategic changes are relatively easy to execute because he has both the mandate and the ideas. In another 300 days, the challenge will be to “produce something more surprising” and achieve cultural compatibility—deepening collaboration, co-design, and figuring out how Tencent’s large-organization culture interacts with the smaller team. “Those challenges and difficulties are greater.”
The cautionary example is 陆奇: he reshuffled the organization aggressively and set the strategy of “stabilize the main channel, search, and win decisively in AI,” after which large numbers of senior executives left Baidu’s core businesses. “Your adjustments did not account for organizational-cultural compatibility.” 曼祺 noted the difference in timing: around 2017, AI could not support a second growth curve in revenue; now that thesis has been partially validated, but uncertainty remains. Some companies say they were profitable in Q2 and Q3 while price-war pressure is heavy, and the path and pace of expansion beyond coding are “not easy to see.” SAP and OpenAI have formed deployment joint ventures with major PE firms worth billions and tens of billions of dollars: “The strongest models do not land in large enterprises automatically; you need many FDEs, or forward-deployed engineers.”

21. The C Paradox: Where Does the Confidence in Staying in the Game Come From?

洪灏’s test for declaring Tencent a latecomer that catches up is straightforward: produce a first-tier model. That leads to what he calls his “longstanding confusion”: everyone says ByteDance’s C team and DeepSeek are certain to stay in the game. C’s team has the best resources, talent, money, and infra, “but it genuinely has not produced the kind of reputation associated with Kimi, DeepSeek, or Qwen.” In coding, it “may not even be at the top of the second tier.” “How can everyone so easily declare who will stay in the game? Where does the evidence come from?”
曼祺’s answer is a memory of short-term historical momentum. ByteDance was China’s strongest technology company over the past 5-8 years, with the best monetization and organizational execution. “That is a very emotional source.” The observable signal is that people leaving ByteDance raise money more easily; content operators from NetEase may be capable but struggle to raise. ByteDance is also one of the few companies to repeatedly build successful products from scratch—Fanqie and Hongguo both reached over 100M DAU. Tencent teams are now bringing in ByteDance talent in large numbers to replace older teams and their operating philosophy.
But ByteDance alumni starting companies face a double edge: their efficient loop “mostly did not come from you; it came from the company’s accumulated experience over many years.” 张一鸣 deliberately attributed product success to the system. “ByteDance will not have a father of Douyin or a father of Toutiao”; even 张楠 cannot be called the person who created Douyin, unlike products such as WeChat Pay that carry a much stronger individual imprint.

22. Coding’s SaaS Anxiety: Seat Utilization, the White-Collar Inflection Point, and Calling the Top Every Year

This year’s competition among the 3 companies is not as “life-or-death” as last year’s chatbot battle. That was a super-entry-point narrative—“the next Douyin, the next WeChat; miss it and it may be gone”—with the Lunar New Year red-packet war as its climax, and it has now been disproved. Coding is more of a productivity tool, and 洪灏 worries it will develop into a SaaS ecosystem: slow growth, lower revenue than C-end products, and no single winner. “Is its imagination space in China simply quite limited?” 曼祺 added that the real upside lies in general-purpose agents supported by coding; the market is the white-collar population, “whose size can be calculated.”
The deployment data is not encouraging. An enterprise may buy 10,000 seats, but the actual employee usage rate is low. “Whether you look at selling by seat or at C-end payments, it is not really that optimistic.” The growth slope from programmers to Pro-C users remains “very low”; it has “not reached an inflection point.” One counterintuitive fact: large numbers of programmers at ByteDance, Tencent, and Alibaba use Doubao and Qwen to write code. 曼祺 admitted, “That is different from what I imagined.” The pattern she had seen was that core users relied on OpenAI and Anthropic, with domestic open-source models filling the value-for-money slot.
The industry shifted before public opinion did. After Anthropic took off with “4.6,” major companies had already begun pivoting into coding by the end of last year; coding simply did not break into the mainstream. Zhipu and Kimi had made it their highest-priority direction since the first half of 2025. The timing remains hard to read: “Since 2023, every autumn people have said we’ve reached the top and the bubble was about to burst—it has happened 3 times.” The industry has “a rhythm of autumnal melancholy”—“not spring melancholy; everyone is highly optimistic in spring.” So far, commercialization has produced only 2 validated categories: coding and video generation.

What an AI Learns to Optimise For as You Train It Harder — Apollo Research How Decagon Runs 90% of Its Agents on Open-Source Models