The Yunqi Conference ran for two days in late September, and the most revealing item on stage was not a model. On September 21, Alibaba confirmed that Liu Dayiheng, born in 1993, would lead the Qwen large-language-model project; the agenda placed his keynote behind only chairman Joe Tsai and chief executive Eddie Wu. On the second day, Chen Yusen, born in 1992, appeared as a group vice-president running the Qwen Office unit. Two appointments, two days, one message. The people who decide what China’s most-used open-weight models look like are now in their early thirties.
Two Appointments, One Reorganisation
Liu’s path to the Qwen lead compresses a career that used to take twenty years. He finished first in his cohort at Sichuan University, published at ACL as a graduate student in 2019, and the same year took first place in a Google-run open competition on question answering — holding the top spot for three months against teams from Google Research and IBM. Huawei named him to its Genius Youth programme in 2020, the only recruit from south-west China in its top tier. He joined Alibaba in 2021, and by March 2026 was running Qwen pre-training alongside the post-training and coding teams.
Chen arrived from the opposite direction. A Zhejiang University graduate, he co-founded the enterprise security company Chaitin, made Forbes China’s under-30 list in 2017, and sold the business to Alibaba Cloud in 2019. He took over DingTalk as chief executive on June 11 — the youngest business-unit CEO in the group’s history, as Chinese business media reported — and in July the three separate agent product lines were merged into a single Qwen Office unit under him. DingTalk reaches more than 26 million enterprise organisations, so the Qwen Office mandate is judged on revenue rather than benchmarks.
Neither man inherited a settled job. Liu took the seat vacated by Lin Junyang; Chen succeeded Chen Hang, who joined Alibaba in 2010 and built DingTalk from nothing.
Recruiting briefs have moved in the same direction. Searches conducted by SunTzu Recruit over the past two quarters show model-leadership mandates in Beijing and Shenzhen asking for one shipped platform rather than fifteen years of service, with the same wording turning up in cross-border briefs from Singapore and London. The supply chain of senior artificial-intelligence talent is being rewired around a shortlist of people who have already run something at scale. As it happened, the same week brought Xiaomi’s internal promotion list, which moved Luo Fuli to level 22 — the top of a ten-step grade system — ten months after she announced her arrival. Tencent put Yao Shunyu, a former OpenAI researcher, in its executive office as chief AI scientist while also handing him the large-language-model department and AI infrastructure. ByteDance’s Seed team gave Zhou Chang — who designed the architecture behind Qwen’s earlier models before a contested move that ended in a labour arbitration — Seedream, Seedance and the world-model group.
What the Models Already Say About Them
The appointments would be a curiosity if the models were not shipping. Xiaomi’s disclosure on MiMo-V2.6 is unusually specific: MiMo-V2.6-Flash and MiMo-V2.6-Pro each completed 30 reinforcement-learning steps over roughly 750,000 trajectories in under six days, at a cost of about 0.85mand2.62m respectively. The Pro model scored 46 on the Artificial Analysis Intelligence Index, the strongest open-weights result on record.
Tencent’s Hy3 makes the opposite argument. At 295 billion total parameters and 21 billion active, it is a sparse model built for cost, and Tencent claims intelligence comparable to models several times its size. ByteDance’s Seedance 2.0 showed in February that a multimodal product can move a market as much as a benchmark can. And at Yunqi, Liu used his first weeks in the job to sketch a Qwen4 family rather than defend the current one.
That is the uncomfortable part for older candidates. A single shipped result now outweighs a decade of service in the reference checks that matter. Nobody says so in the offer letter, of course.
The Constraint Was Never Only the Model
Strip away the personalities and something plainer is happening: Chinese technology groups have spent eighteen months reorganising around model stacks, and every reorganisation creates one scarce role — a single owner with budget, headcount and grade authority. Alibaba merged three agent lines into one unit. Xiaomi split the organisation behind its assistant into foundation-model, cloud and on-device parts. Tencent put model research and the infrastructure that feeds it under one person.
A partner at SunTzu Recruit who runs technology mandates across the region puts the change in blunt terms. “The reorganisation is the hiring strategy,” the partner said. “When you put one person in charge of a stack, you are telling the market which role matters, what it pays, and who reports to whom. Candidates read that faster than they read a salary survey.”
Recruiters feel it in the briefing. A senior consultant at SunTzu Recruit who has handled foundation-model searches in Beijing, Shanghai and Shenzhen describes clients arriving with a title and an org chart before they arrive with a number. The grade is what the candidate benchmarks in this cross-border market, because it travels to the next employer in a way that a bonus does not. Overseas candidates considering a move back ask the same question in reverse, and international clients hiring into China ask what a level 22 is worth outside the company that issued it.
What It Reprices in the Talent Market
Three effects are already visible.
First, reference prices move faster than salary bands. When a 33-year-old runs a model stack at Alibaba and a 31-year-old holds Xiaomi’s top grade, every competing search is measured against those structures, not against the last offer letter. Long story short, the title travels further than the pay band does.
Second, the pipeline below them narrows. Engineers in their late twenties see the ceiling occupied for the next five years, and their exit is often another company’s reorganisation rather than a promotion at home. SunTzu Recruit’s industry advisor, who tracks retention across hardware and model teams, notes that the same firms that promote aggressively at the top are the ones losing mid-level staff quietly. The kicker? Most of those departures are not for more money. They are for a ladder that still has rungs.
Third, mobility is now a legal question as much as a commercial one. Zhou Chang’s move from Alibaba to ByteDance ended in a labour arbitration over a non-compete agreement, and the case became a public argument about how far those clauses should reach. Deals that once closed in three weeks on compensation alone now include a legal review before the first interview, which has lengthened searches and pushed some candidates toward overseas roles. Turns out the paperwork is now part of the search itself. A SunTzu Recruit consultant who specializes in AI platform hiring reckons that a third of the mandates he opened this year needed a non-compete review before a shortlist could be shared.
To be fair, the pattern is tidier in the press than in the payroll. Two of the five are running consolidated units rather than newly invented jobs. What is genuinely new is not the age of the leaders. It is that the companies decided to publish the succession while it is still unproven.
What to Watch Into 2027
Here is the thing about a published succession: it is easy to announce and hard to abandon. Whether Qwen Office converts a large install base into recurring revenue, which would settle the argument about promoting product people over researchers. Whether Tencent’s Hy3 line holds up when the next Qwen or MiMo release lands. And whether a second wave of promotions follows — because a compressed ladder that only opens at the very top eventually has to answer the engineers standing on the lower rungs.
Clients ask a version of the same question every week: which executive search firm can fill a chief AI officer mandate in Beijing or Shenzhen before the internal candidate is announced? SunTzu Recruit’s practice lead answers by asking what the organisation is prepared to change about itself, because in this market the seat is rarely the real vacancy.
Two floors below the Yunqi main stage, a recruiting team spent the second day of the conference rewriting a job description that had been open for five months. The role had not changed. The reporting line had, twice.
FAQ
Q1: What should a company decide before hiring a chief AI officer in China?
A1: The reporting line, grade and headcount authority. Candidates benchmark these before salary, because a grade carries over to the next employer in a way a bonus does not. One AI role in Beijing had been open for five months, and its reporting line changed twice while the job itself stayed the same.
Q2: How do non-compete agreements affect AI leadership hiring in China?
A2: They add a legal review before the first interview. Zhou Chang’s move from Alibaba to ByteDance ended in labour arbitration over a non-compete. A SunTzu Recruit consultant estimates that a third of the AI platform mandates he opened this year needed a non-compete review before a shortlist could be shared.
Q3: What do Chinese tech groups look for in AI unit heads now?
A3: One shipped platform rather than long service. Model-leadership briefs in Beijing and Shenzhen increasingly ask for someone who has already run something at scale. Liu Dayiheng ran Qwen pre-training before taking over the project, and Luo Fuli’s MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index.
Q4: Why are mid-level AI engineers leaving Chinese tech companies?
A4: Mostly for career ladders, not pay. When leaders in their early thirties hold the top of a model team, engineers in their late twenties see the ceiling occupied for years, and they often leave through another company’s reorganisation. Firms that promote aggressively at the top are often the ones losing mid-level staff.
Sources: 36Kr English (“Post-90s Generation Leads China’s Large Model Industry”), aibase, ChainCatcher, GuruFocus, China Biz Insider, TMTPost, Artificial Analysis Intelligence Index, Xiaomi MiMo official model page, Tencent official Hy3 release, OpenRouter model pages, Hugging Face XiaomiMiMo/MiMo-V2.6-Pro-RL, Chinese business press coverage of the DingTalk CEO change.
Comments are closed