Luo Fuli

Luo Fuli (Chinese: 罗福莉; also written Fuli Luo) is a Chinese AI researcher who leads the MiMo large-model team at Xiaomi. Before Xiaomi she worked at Alibaba's DAMO Academy, where she was first author of the cross-lingual pre-trained model VECO, and then at the hedge fund High-Flyer and its AI company DeepSeek. Her name appears on the author lists of DeepSeek papers including DeepSeek-V2, DeepSeek-V3 and DeepSeek-R1.^{[1][3][9][12][13]} She confirmed that she had joined Xiaomi MiMo on 12 November 2025. Since then she has been listed as corresponding author on Xiaomi's MiMo-V2-Flash and MiMo-V2.6 technical reports.^{[1][4][16][18]}

Luo first drew wide attention in China in 2019. That year, while she was a master's student at Peking University, her name appeared on eight papers at the annual meeting of the Association for Computational Linguistics (ACL).^[1][7] Chinese media later gave her the label "AI genius girl", which she has publicly rejected.^[4][6] Reuters and the South China Morning Post report that she was born in 1995.^[1][2]

At a glance

Education and early research

Jiemian News and other Chinese outlets report that Luo studied computer science as an undergraduate at Beijing Normal University. She then did a master's in computational linguistics at Peking University's Institute of Computational Linguistics.^[3][4] A January 2026 Peking University profile calls her an alumna of the School of Electronics Engineering and Computer Science.^[6] Her student papers list the Key Laboratory of Computational Linguistics (Ministry of Education) at Peking University as her affiliation.^[8]

Her first-author papers from this period covered word sense disambiguation, text style transfer, controllable text generation and pun generation. The ACL Anthology lists eight papers with her as a co-author at the 57th Annual Meeting of the ACL in 2019. She is first author on two of them: "Learning to Control the Fine-grained Sentiment for Story Ending Generation" and "Towards Fine-grained Text Sentiment Transfer."^[7] This is the "eight ACL papers" episode that the press often mentions.^[1][3]

Alibaba DAMO Academy

After graduating, Luo joined Alibaba's research institute, DAMO Academy.^[1] There she worked on pre-trained language models. Chinese reports say she led the development of the multilingual model VECO and pushed forward the open-sourcing of AliceMind, Alibaba's collection of pre-trained language models.^[3][5] The VECO paper, first posted to arXiv in October 2020 and accepted as an ACL 2021 long paper, lists Luo and Wei Wang as equal-contribution first authors, both at Alibaba Group.^[9] The AliceMind GitHub repository lists VECO among its models.^[10] She was also first author of "Rethinking Denoised Auto-Encoding in Language Pre-Training" (EMNLP 2021). She co-authored papers on model pruning, cross-lingual fine-tuning and prompt tuning, some of which list her affiliation as Alibaba Group.^[7][11]

High-Flyer and DeepSeek

According to the South China Morning Post, Luo joined the quantitative hedge fund High-Flyer in 2022 and later moved to DeepSeek, the AI company High-Flyer owns, as a deep-learning researcher. Both firms were founded by Liang Wenfeng. The Post says she contributed to DeepSeek-V2, which was released in May 2024.^[1] Chinese business media have described her as one of DeepSeek-V2's key developers.^[3][5]

She appears on the author lists of several DeepSeek papers. DeepSeek's large technical reports use long team author lists, so being listed does not show which part of the work a person did.

In her 21 September 2026 post about MiMo-V2.6, Luo described DeepSeek R1 as a project "which I was partly involved in."^[19]

Move to Xiaomi

In late December 2024, Chinese media reported that Xiaomi founder Lei Jun had personally tried to recruit Luo with an annual salary in the tens of millions of yuan to lead Xiaomi's large-model team. The 21st Century Business Herald attributed the report to the Securities Times.^[5] Neither Xiaomi nor Luo confirmed the offer.^[1] On 18 February 2025, Luo wrote on WeChat Moments asking the internet to leave her alone to work. She said she had said years earlier that she was not a "genius girl". She also criticized self-media accounts that had contacted her family, friends and former teachers.^[4]

On 12 November 2025 she said in a WeChat post that she had joined Xiaomi MiMo. The South China Morning Post translated her post as: "Intelligence will step beyond language into the physical world. I'm at Xiaomi MiMo, with a group of creative, talented and sincere researchers to build a future like that."^[1] In the paper's second arXiv version (21 October 2025), Luo is listed as an independent researcher and co-corresponding author on "Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers" (Rollout Routing Replay). That paper was first posted to arXiv in October 2025, and its other authors are from Xiaomi's LLM-Core team and Peking University.^[15]

Leading Xiaomi MiMo

Luo first appeared in public as head of the MiMo team on 17 December 2025, at Xiaomi's "Human x Car x Home" ecosystem partner conference. There she presented MiMo-V2-Flash, a 309-billion-parameter mixture-of-experts model with 15 billion active parameters. Jiemian News reported that she named three features: strong coding and tool-calling ability, a model structure designed around inference efficiency, and a new post-training paradigm for scaling reinforcement learning.^[3] The MiMo-V2-Flash technical report, published by "LLM-Core Xiaomi", lists Luo as its corresponding author.^[16] The report also introduced Multi-Teacher On-Policy Distillation (MOPD). A June 2026 paper on MOPD, with Luo and Sui Zhifang as co-corresponding authors, says the method was used in MiMo-V2-Flash post-training.^[17]

In March 2026, Reuters reported that "MiMo is led by ex-DeepSeek researcher and PKU graduate Luo Fuli". The report came a day after Xiaomi unveiled its MiMo-V2-Pro model. At a company event in Beijing that day, Lei Jun said the MiMo Core Team behind the models had an average age of 25, that about 60% of its members came from Tsinghua or Peking University, and that 55% held doctorates.^[2][23][24] A July 2026 paper on inference optimization for the MiMo-V2.5 series also names her as corresponding author.^[22]

On 17 September 2026, Luo wrote that after open-sourcing MiMo-V2.5 in April, Xiaomi had spent close to half a year studying how far reinforcement learning could scale. IT Home reported that she also shared a public live-stream of the MiMo-V2.6 RL run, whose spending had passed $1.25 million at that point.^[20] On 21 September 2026 she published a long post on X, "MiMo-V2.6: The Hard Road to Scaling Up RL." In it she called MiMo-V2.6 "very likely one of the largest single RL runs, by compute, that any open-source model team has undertaken to date." She said the team had dedicated "a team of several dozen people to one goal over an extended period: scaling up RL."^[19] She described MixRL and multi-teacher distillation as complementary rather than "competing choices." In her account, verifiable tasks of moderate difficulty such as code went into one mixed RL run. Hard-to-verify or very long-horizon tasks were trained separately and merged through MOPD.^[19] The MiMo-V2.6 technical report lists her as its only corresponding author.^[18]

Xiaomi papers with Luo as corresponding author (selection)

Public profile

Luo has pushed back against the "genius" label. In the Peking University profile she said, "I'm not a genius," and credited "curiosity about fundamental problems, solid execution, and a bit of stubborn resilience" instead.^[6] In the same profile she said that she thinks progress toward real intelligence will require learning from observation and interaction with the physical world, not only from text.^[6] She uses the X handle @_LuoFuli for English-language posts about MiMo.^[19]

See also

• Xiaomi AI
• Xiaomi MiMo-V2.6
• Xiaomi MiMo-V2.5
• DeepSeek
• Liang Wenfeng

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Ben Jiang, "Chinese AI prodigy Luo Fuli joins Xiaomi as industry competition for talent heats up," South China Morning Post, 12 November 2025. scmp.com/...aomi-industry-competition-talent-heats
2. ^1 ^2 ^3"Xiaomi to invest at least $8.7 billion in AI over next three years, CEO says," Reuters, 19 March 2026. reuters.com/...ext-three-years-ceo-says-2026-03-19
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Song Jianan, "加入小米一个多月后，'AI才女'罗福莉完成首秀" (A month after joining Xiaomi, Luo Fuli makes her debut), Jiemian News, 17 December 2025. jiemian.com/...13773562
4. ^1 ^2 ^3 ^4 ^5 ^6"'AI才女'罗福莉官宣加入小米，曾传雷军千万年薪挖角" (Luo Fuli announces she has joined Xiaomi), Jiemian News via Sina Finance, 12 November 2025. finance.sina.com.cn/...doc-infxcqie4891668.shtml
5. ^1 ^2 ^3"千万年薪！雷军突然宣布" (Lei Jun's recruitment offer report), 21st Century Business Herald (21jingji.com), 29 December 2024. 21jingji.com/...286d48229f87810b10c897a3931fc7be
6. ^1 ^2 ^3 ^4Nicole Angriawan, "How Luo Fuli Keeps an Earthy Touch as she Soars Through the AI World," Peking University News, 15 January 2026. newsen.pku.edu.cn/...15385
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Fuli Luo," ACL Anthology author page. aclanthology.org/...fuli-luo
8. ^1 ^2Fuli Luo, Tianyu Liu, Qiaolin Xia, Baobao Chang and Zhifang Sui, "Incorporating Glosses into Neural Word Sense Disambiguation," ACL 2018, arXiv:1805.08028. arxiv.org/...1805.08028
9. ^1 ^2Fuli Luo, Wei Wang et al., "VECO: Variable and Flexible Cross-lingual Pre-training for Language Understanding and Generation," ACL 2021, arXiv:2010.16046. arxiv.org/...2010.16046
10. ^Alibaba, "AliceMind: ALIbaba's Collection of Encoder-decoders from MinD," GitHub. github.com/...AliceMind
11. ^Runxin Xu, Fuli Luo et al., "From Dense to Sparse: Contrastive Pruning for Better Pre-trained Language Model Compression," arXiv:2112.07198. arxiv.org/...2112.07198
12. ^1 ^2DeepSeek-AI, "DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model," arXiv:2405.04434, May 2024. arxiv.org/...2405.04434
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7arXiv author listings for DeepSeek LLM (arXiv:2401.02954), DeepSeek-Coder (arXiv:2401.14196), DeepSeek-Coder-V2 (arXiv:2406.11931), DeepSeek-Prover-V1.5 (arXiv:2408.08152), DeepSeek-V3 Technical Report (arXiv:2412.19437) and DeepSeek-R1 (arXiv:2501.12948). arxiv.org/...2501.12948
14. ^Damai Dai et al., "DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models," arXiv:2401.06066. arxiv.org/...2401.06066
15. ^1 ^2Wenhan Ma, Hailin Zhang, Liang Zhao, Yifan Song, Yudong Wang, Zhifang Sui and Fuli Luo, "Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers," arXiv:2510.11370, October 2025. arxiv.org/...2510.11370
16. ^1 ^2 ^3LLM-Core Xiaomi, "MiMo-V2-Flash Technical Report," arXiv:2601.02780, January 2026. arxiv.org/...2601.02780
17. ^1 ^2Wenhan Ma et al., "MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training," arXiv:2606.30406, June 2026. arxiv.org/...2606.30406
18. ^1 ^2 ^3LLM-Core Xiaomi, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," technical report, September 2026. huggingface.co/...MiMo_V2_6_technical_report.pdf
19. ^1 ^2 ^3 ^4 ^5Fuli Luo (@_LuoFuli), "MiMo-V2.6: The Hard Road to Scaling Up RL," X, 21 September 2026. x.com/...2102162926802968749
20. ^"小米直播训练 MiMo-V2.6 模型，罗福莉称沉寂半年钻研一件事" (Xiaomi live-streams MiMo-V2.6 training), IT Home, 17 September 2026. ithome.com/...555
21. ^Fuli Luo et al., "A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer," IJCAI 2019, arXiv:1905.10060. arxiv.org/...1905.10060
22. ^1 ^2Xiaomi MiMo Team et al., "Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit," arXiv:2607.13095, July 2026. arxiv.org/...2607.13095
23. ^Hou Ruining, "雷军称未来三年小米AI投入将超600亿，首款手机龙虾产品已启动封测" (Lei Jun says Xiaomi's AI investment will exceed 60 billion yuan over three years), Jiemian via Securities Times (stcn.com), 20 March 2026. stcn.com/...3687665
24. ^"小米AI大模型三连发！雷军：未来三年将在AI领域投入超600亿元" (Xiaomi releases three AI models; Lei Jun: more than 60 billion yuan for AI over three years), China Securities Journal via China Daily (chinadaily.com.cn), 20 March 2026. cn.chinadaily.com.cn/...WS69bcf34ea310942cc49a4239