In late 2024, Chinese outlets reported that Xiaomi’s founder had offered a ten-million-yuan salary to a then 29-year-old researcher to lead the company’s large-model team [8]. The number travelled further than her papers did. Most of the coverage was about the number.
The useful question is not why one company paid it. It is why the pool of people who could fill that chair is small enough that a number like that makes sense. That question has an answer, and it is visible in a publication record.
What the record contains
| Year | Work | Position |
|---|---|---|
| 2018 | Word sense disambiguation with glosses, ACL | first author, master’s era |
| 2020 | VECO, cross-lingual pre-training | first author, Alibaba |
| 2021 | PLUG, a 27B Chinese pre-trained model, and AliceMind | lead developer per reports; open-source driver |
| 2024 | DeepSeek-V2, the MoE model | named contributor |
| 2026 | MiMo-V2.6 and the MiMo inference line | team lead |
The first-author work starts in 2018 at Peking University and turns to pre-training by 2020 [5]. Five weeks after she joined Xiaomi she presented the team’s first model in public, a mixture-of-experts build at 309B total and 15B active parameters [7], and the MiMo papers from 2026 carry her name alongside the team she leads [4]. VECO is unambiguously hers: first author, Alibaba Group, ACL 2021 long paper [1]. PLUG and the AliceMind release are the production half of the same period [2].
By the time the field reorganised around pre-training at scale, she had a paper in it and a 27B model shipped. That ordering matters more than any single result.
One part of the record is weaker than it looks. Being listed on DeepSeek-V3 and R1 says little on its own, because both papers carry more than 150 authors. V2 is the line that matters, not because it is bigger, but because it is the model that made mixture-of-experts an economic argument rather than a research one [3].
Four things that made the record transferable
She specialised before the field did. Pre-training at 27B scale was a narrow skill in 2020. A first-author paper on multilingual pre-training, two years before every company wanted one, is a bet that paid.
She has both halves of the job. First-author papers at ACL and EMNLP, and production work at scale. Most researchers have the first half. Labs building a frontier model from nothing need a person who can do the research and ship the training run, because at that stage one person often does both.
She has a legible credential. A recruiter cannot evaluate a paper on contrastive pre-training. A recruiter can read “core contributor to DeepSeek-V2” and understand what it means. Legibility is not the same as depth, and in a hiring window measured in months it is what moves a decision.
The network was already in place. The MiMo papers connect to her former Peking University group, which appears as corresponding authors on work published before she had officially joined Xiaomi [5]. The offer landed on a relationship that already existed.
What this says about hiring, not about her
Citations are a lagging indicator. Her count passed 11,000 in late 2025, with about 8,000 of those added in a single year [5]. Nobody hired her for the count. The count followed the shipped models.
The currency is a training run, not a publication. Every property above that transferred came from work that consumed GPU time and shipped weights: PLUG, AliceMind, V2. The papers described them; they did not cause the offers.
Specialise where the field is going, not where it is. Pre-training was unglamorous in 2020 and the centre of everything by 2023. Four years is enough time for a bet like that to become a hiring advantage, and too short for a plan.
And the supply side explains the price better than the demand side does. The number of people who have led a frontier-scale training run is small. A paper with 150 authors does not represent 150 people who could run one. Xiaomi had no frontier team in late 2024 and a hardware ecosystem that needed one [6]. When the buyers are few and the sellers are fewer, one hire becomes a strategic purchase.
What this does not establish
The salary is reported, not confirmed. Xiaomi has not published a figure, and no primary document states one.
It is also a single case. A profile of one successful path cannot tell you what the other people in the pool did, or how many were passed over for reasons the record does not show.
One detail is worth flagging because it is usually reported as fact and is not. On DeepSeek’s Nature paper, she is listed as an independent researcher in Beijing at a time when she was reported to be at Xiaomi [5]. The plausible reading is a non-compete. That reading is an inference, and the note stays an inference.
References
[1] VECO: Variable and Flexible Cross-lingual Pre-training for Language Understanding and Generation: the ACL 2021 paper, first-authored by Luo, that anchors the pre-training claim.
[2] alibaba/AliceMind: the open-source model family that includes PLUG (27B), anchoring the production-scale claim.
[3] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model: the MoE report, where her name appears among the contributors.
[4] MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement: the Xiaomi paper carrying both her name and the team she now leads.
[5] 罗福莉C位亮相小米，离职DeepSeek后首次官宣 (量子位, via 智源社区): the education path, the Alibaba and DeepSeek roles, the citation count, and the independent-researcher detail.
[6] 前DeepSeek研究员罗福莉已加入小米 (澎湃新闻): the 12 November 2025 join announcement and her own statement about the move.
[7] 加入小米一个多月后，“AI才女"罗福莉完成首秀 (界面新闻): the first public appearance and the MiMo-V2-Flash numbers.
[8] 雷军千万年薪聘用的95后女生罗福莉，履历曝光 (每日经济新闻): the salary reporting, stated here as a report rather than a confirmed figure.