Yang Zhilin

Yang Zhilin (Chinese: 杨植麟; pinyin: Yáng Zhílín) is the co-founder and chief executive officer of Moonshot AI (月之暗面), the Beijing startup that develops the Kimi chatbot and the Kimi family of large language models, including the trillion-parameter Kimi K2 and the roughly 2.8-trillion-parameter Kimi K3. ^{[1][2][5][36]} He founded Moonshot AI in 2023 and is one of the most academically credentialed founders in Chinese AI: he holds a PhD from Carnegie Mellon University and was an equal-contribution first author of the influential natural language processing papers Transformer-XL and XLNet. ^[3][4][22] In July 2026 Chinese media reported that Moonshot had closed an F round of more than 3.5 billion US dollars at a post-money valuation of about 35 billion dollars, and business registration records filed that month list Yang as the company's chairman and manager. ^[31][32]

Moonshot AI is often grouped with a small set of well funded Chinese startups that some investors call the country's "AI tigers," and it has attracted backing from Alibaba, Tencent, Meituan, and other large technology firms. ^[5][6][17] In his pinyin name the family name is Yang and the given name is Zhilin. English language coverage and his own research papers usually render the name as Zhilin Yang, and he goes by Kimi in English, the name the company gave its chatbot. ^[3][6]

Who is Yang Zhilin?

Yang Zhilin is a Chinese artificial intelligence researcher and entrepreneur, born in Shantou in Guangdong province. He is best known in two roles: as a researcher who co-authored XLNet and Transformer-XL during his doctorate at Carnegie Mellon, and as the founder and CEO of Moonshot AI, the company behind the Kimi assistant. ^[1][3][5] As of 2026 he leads Moonshot AI, where he has positioned long-context and agentic models as the company's route toward artificial general intelligence. ^[9][13][16]

Early life and education

Yang was born in Shantou, in China's Guangdong province. Most English language sources give his birth year as 1992, while some Chinese profiles list 1993. ^[1][2] As a teenager he was drawn to music and at one point imagined a future as a rock musician, an interest that later shaped how he and his co-founders named their company and its meeting rooms. ^[1][6][7]

He studied computer science at Tsinghua University in Beijing and received his bachelor's degree in 2015. His undergraduate adviser was Tang Jie, the Tsinghua professor who later co-founded Zhipu AI (now branded Z.ai). ^[6][22] During his undergraduate years Yang played drums and wrote songs for a campus rock band called Splay, whose name echoes the splay tree data structure. ^[1][6]

After Tsinghua, Yang moved to the United States for doctoral study in the School of Computer Science at Carnegie Mellon University, where his thesis was published through the Language Technologies Institute. His advisors were Ruslan Salakhutdinov, a machine learning researcher who directed AI research at Apple, and William W. Cohen, a scientist at Google DeepMind; the two co-chaired his thesis committee, which also included Graham Neubig of Carnegie Mellon and Jason Weston of Facebook AI Research. ^[6][22][23] Yang completed the PhD in 2019, in about four years rather than the longer span that is typical in the United States, with a thesis titled "Advances in Generative Feature Learning." ^[6][23] During his studies he also worked at industry research labs: at Google Brain with Quoc V. Le, and at Facebook AI Research (now Meta AI) with Jason Weston. ^[7][22]

What did Yang Zhilin research before Moonshot AI?

Yang's academic work centers on representation learning and language modeling. His thesis framed this as "generative feature learning": using generative modeling of unlabeled data to improve performance on target tasks, both through unsupervised pretraining and through semi-supervised learning. ^[23] During his years in the United States he co-authored papers with researchers including the Turing Award winners Yoshua Bengio and Yann LeCun, and he co-authored the multi-hop question answering dataset HotpotQA. ^[6][7][22]

Transformer-XL

Transformer-XL, posted in January 2019 and published at ACL 2019, lists Zihang Dai and Yang as equal-contribution first authors, followed by Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov. ^[22][24] It addressed a limitation of the original Transformer for language modeling. Standard Transformers process text in fixed length segments and cannot easily carry information across segment boundaries, which caps the amount of context a model can use. Transformer-XL added a segment level recurrence mechanism and a relative positional encoding scheme so that hidden states from earlier segments could be reused, letting the model attend to a longer stretch of preceding text. ^[4][8][24] The design improved language modeling results and became a building block for later work on long sequences.

XLNet

XLNet, published in 2019, was authored by Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le, with Yang and Dai marked as equal contributors; it was presented as an oral paper at NeurIPS 2019. ^[3][22] It proposed a training objective called permutation language modeling. Rather than masking tokens in the style of BERT, XLNet maximizes the expected likelihood over many possible orderings of the input sequence, so that each position can learn from tokens on both its left and its right while keeping an autoregressive formulation. ^[3] Because it does not corrupt the input with mask symbols, XLNet avoids the mismatch between pretraining and fine tuning that affected masked models, and it used Transformer-XL as its backbone to handle longer context. ^[3] On its release XLNet reported gains over BERT across roughly twenty language understanding tasks, including question answering, natural language inference, sentiment analysis, and document ranking. ^[3] The two papers were among the more heavily cited NLP works of that period, and XLNet alone has been cited more than ten thousand times. ^[6]

The research connects directly to the work Yang would later pursue in industry. Both Transformer-XL and XLNet are concerned with extending how much context a model can use, a theme that runs through Moonshot AI's emphasis on long-context systems. ^[9][23]

Work with Tang Jie's group and GLM

After his doctorate Yang continued to publish with his former undergraduate adviser's group at Tsinghua. He is one of seven authors, alongside Tang Jie, of "GLM: General Language Model Pretraining with Autoregressive Blank Infilling" (Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang), first posted in March 2021 and published at ACL 2022. ^[22][25] Tang's group later used the GLM pretraining method as the basis of its GLM model line. His homepage lists other 2021 and 2022 papers written with Tang, including P-Tuning v2, FewNLU, FlipDA, and an inverse-prompting paper at KDD 2021, and notes the September 2022 release of the multilingual code generation model CodeGeeX. ^[22] As a result, the founder of Moonshot AI and a co-founder of Zhipu AI appear together on the original GLM paper.

Other ventures before Moonshot AI

While still a student, Yang co-founded Recurrent AI in 2016, a company that applied language technology to sales and customer conversation analysis. ^[1] In China he also took part in large model efforts during the early years of the field. In 2020 he worked with Huawei on an early version of its PanGu model, and in 2021 he led a team working on the Wu Dao large model at the Beijing Academy of Artificial Intelligence. ^[7] A dispute later arose between Yang and five early investors in Recurrent AI, led by GSR Ventures China, which claimed he had an "uncleared liability" when he started Moonshot AI and took the case to the Hong Kong International Arbitration Centre; Yang told the South China Morning Post that Recurrent AI's board had already approved his exit. ^[26]

What is Moonshot AI?

Moonshot AI is the Beijing foundation-model company Yang co-founded in March 2023, together with former Tsinghua schoolmates Zhou Xinyu, Wu Yuxin, and Zhang Yutao; Zhou had played with Yang in Splay. ^[6][7] The company's Chinese name, 月之暗面, translates as "the dark side of the moon," a reference to the Pink Floyd album, Yang's favorite, and the company launched around the album's 50th anniversary. Meeting rooms in its offices carry the names of rock bands such as the Rolling Stones and Led Zeppelin, and one is called Splay. ^[6][27] Yang serves as chief executive. ^[2][5] After the company converted into a joint stock limited company on 29 July 2026, registration records listed him as chairman and manager, with Zhou Xinyu moving from manager to director. ^[32]

Moonshot AI built its public profile around long-context chat assistants under the Kimi brand. In October 2023 it released its first assistant, which could take in inputs of up to about 200,000 Chinese characters in a single conversation, a figure that stood out at the time. ^[5][6] In March 2024 the company said Kimi could handle roughly 2 million Chinese characters of context, and the product drew heavy traffic among Chinese users. ^[6][9] Yang has framed long context as central to the company's strategy, arguing that the ability to process very long inputs without loss is a foundation for useful and personalized AI. ^[9] In a February 2024 interview with the Chinese publication Overseas Unicorn, as translated by ChinaTalk, he put it in stark terms: "If you have a billion-token context length, then the problems we face today would no longer be problems." ^[21]

Kimi k1.5

In January 2025 Moonshot released Kimi k1.5, a reasoning oriented model trained with reinforcement learning. The company reported that it matched OpenAI's o1 on several mathematics, coding, and multimodal reasoning benchmarks, citing scores such as 77.5 on AIME and 96.2 on MATH 500. ^[10] The accompanying technical report described a reinforcement learning approach built on a variant of online mirror descent and stated that it reached these results without relying on Monte Carlo tree search, value functions, or process reward models. ^[10] The report presented this as a simple, effective reinforcement learning framework for long chain of thought reasoning. ^[10]

What is Kimi K2?

In July 2025 Moonshot released Kimi K2, a mixture of experts model with 1 trillion total parameters and 32 billion parameters activated per token, pre-trained on 15.5 trillion tokens with the Muon optimizer and supporting a context window of 128,000 tokens. ^[11][12][33] Kimi K2 was tuned for agentic use, meaning tool calling and multi step task execution, and Moonshot released the code and weights under a modified MIT license in base and instruction tuned versions. ^[11][12][33] Reviewers reported that the instruction tuned model performed strongly relative to other open weight systems on benchmarks for coding, mathematics, reasoning, and tool use. ^[11][12] The company extended the family with a reasoning focused release, Kimi K2 Thinking, in November 2025. ^[34]

Kimi in 2026

Moonshot continued to ship rapidly in 2026. In January 2026 it released Kimi K2.5, a natively multimodal model that understands text, images, and video, alongside a coding agent. ^[18][19] In March 2026 Yang gave a keynote at NVIDIA's GTC conference, where he described the technical roadmap behind K2.5 and summarized Kimi's development as gains along three dimensions: token efficiency, long context, and agent swarms. ^[13] He argued that scaling was "no longer simply about resource accumulation" and predicted that intelligence would evolve from single agents toward dynamically generated swarms. ^[13]

On 20 April 2026 Moonshot released Kimi K2.6. Its model card reported a score of 58.6 on SWE-Bench Pro, against 57.7 for GPT-5.4 at xhigh reasoning effort in the same Moonshot table. ^[20] A coding-specialized Kimi K2.7 Code followed in June 2026. ^[35] On 16 July 2026 Moonshot announced Kimi K3, making it available that day on Kimi, Kimi Work, Kimi Code, and the Kimi API and promising the full weights by 27 July. ^[30] The published model card describes a sparse mixture of experts model with about 2.8 trillion total parameters and 104 billion activated per token, and the weights are distributed on GitHub and Hugging Face under a custom Kimi K3 license. ^[36]

How is Moonshot AI funded?

Moonshot AI raised capital quickly after its founding. According to PitchBook data cited by TechCrunch, it first raised about 200 million US dollars from HongShan (formerly Sequoia China) and ZhenFund at a valuation of about 300 million dollars; the South China Morning Post put its initial funding at 60 million dollars. ^[7][27] In February 2024 the company raised more than 1 billion dollars in a round reported to be led by Alibaba, which valued Moonshot at about 2.5 billion dollars. ^[27] In August 2024 it raised about 300 million dollars in a round that included Tencent, lifting the valuation to roughly 3.3 billion dollars. ^[7][14]

Funding accelerated into 2026. In a round reported by the Chinese outlet LatePost at the turn of the year, Moonshot raised 500 million dollars at a 4.3 billion dollar valuation, led by IDG Capital with Alibaba and Tencent participating; LatePost also cited an internal letter from Yang saying the company held more than 10 billion yuan in cash. ^[28] In February 2026 the South China Morning Post reported that existing investors including Alibaba, Tencent, Andon, and 5Y Capital had jointly led a new round of at least 700 million dollars that could value the company at up to 12 billion dollars; Bloomberg reported a 10 billion dollar valuation target. ^[15][29] In May 2026 TechCrunch reported that Moonshot had raised about 2 billion dollars at a valuation of roughly 20 billion dollars, in a round led by Long-Z Investments, the venture arm of the food delivery company Meituan, with participation from Tsinghua Capital, China Mobile, and CPE Yuanfeng. ^[17] The same report said the company's annual recurring revenue topped 200 million dollars in April 2026, driven by paid subscriptions and API usage. ^[17]

On 29 July 2026 the Chinese outlet The Paper reported that Moonshot had closed its F round early, raising more than 3.5 billion dollars at a post-money valuation of about 35 billion dollars, and that a G round, described as its pre-IPO round, had started ahead of schedule at a pre-money valuation of 50 billion dollars. ^[31]

The table below summarizes the publicly reported funding milestones, attributed to the sources that disclosed them.

Where does Moonshot AI rank among Chinese AI labs?

Moonshot AI is frequently named among a handful of Chinese large model startups that investors and the press call the "AI tigers," a group that also includes companies such as Zhipu AI, MiniMax, Baichuan, and StepFun. ^[5][6] Within that cohort Moonshot became known for its consumer facing Kimi assistant and, after the open weight release of Kimi K2, for its standing among open large language models. ^[11][12] The company's heavy backing from Alibaba, Tencent, and later Meituan places it within a broader pattern of artificial intelligence development in China in which large technology firms invest in independent model builders. ^[5][6][17]

What are Yang Zhilin's views on AGI and scaling?

Yang describes Moonshot AI as a company aimed at artificial general intelligence, and he ties that goal to long context. He has argued that, viewed broadly, many AI problems can be cast as problems of using longer and richer context, and that progress in model architectures has tracked increases in the effective context a model can handle. ^[9][21] In the same 2024 interview he said he hoped Moonshot could become "a company that combines OpenAI's technology idealism with the business philosophy of ByteDance." ^[21]

On the question of whether large model gains will continue, Yang has said that the scaling laws remain valid but that the focus is shifting. In a November 2024 interview tied to the launch of the company's K0-math model, he said that scaling does not stop but proceeds through different approaches, with attention moving toward improving reasoning through reinforcement learning rather than only adding computation. ^[16] He also said there was still room left in pretraining, perhaps for around half a generation to a full generation of models, and that by the measure of distance to AGI the field remained at an early stage. ^[16] In his GTC talk in March 2026 he continued to stress long context and agent swarms as directions for reaching more general intelligence. ^[13]

Recognition

Yang is widely covered in Chinese and international media as a leading figure among the younger generation of Chinese AI founders, and Moonshot AI's rapid rise to a multi billion dollar valuation has been reported by outlets including the South China Morning Post, Bloomberg, TechCrunch, and The Wire China. ^{[5][6][15][17]} His earlier academic work on XLNet and Transformer-XL is frequently noted in these profiles as the basis for his reputation in the field; Salakhutdinov described him to The Wire China as an "absolutely brilliant student." ^[6]

Facts

References

1. ^1 ^2 ^3 ^4 ^5 ^6"Yang Zhilin." Wikipedia. en.wikipedia.org/...Yang_Zhilin
2. ^1 ^2 ^3"杨植麟." Baidu Baike. baike.baidu.com/...24462023
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7Yang, Zhilin; Dai, Zihang; Yang, Yiming; Carbonell, Jaime; Salakhutdinov, Ruslan; Le, Quoc V. "XLNet: Generalized Autoregressive Pretraining for Language Understanding." arXiv:1906.08237, 2019. arxiv.org/...1906.08237
4. ^1 ^2zihangdai. "XLNet: Generalized Autoregressive Pretraining for Language Understanding." GitHub. github.com/...xlnet
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Moonshot AI." Wikipedia. en.wikipedia.org/...Moonshot_AI
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Berman, Noah. "Who Is Moonshot AI?" The Wire China, 28 September 2025. thewirechina.com/...who-is-moonshot-ai
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Jiang, Ben. "Meet Yang Zhilin: Moonshot AI founder builds business in the mould of ByteDance, OpenAI." South China Morning Post, 7 August 2024, via Yahoo Finance. finance.yahoo.com/...-zhilin-moonshot-ai-093000045
8. ^"XLNet." Hugging Face Transformers documentation. huggingface.co/...xlnet
9. ^1 ^2 ^3 ^4 ^5"Moonshot AI: Betting Big on Long-Context, Confronting the Challenges of Scale and Reliability." Center for Data Innovation, January 2025. datainnovation.org/...ges-of-scale-and-reliability
10. ^1 ^2 ^3Kimi Team. "Kimi k1.5: Scaling Reinforcement Learning with LLMs." arXiv:2501.12599, 22 January 2025. arxiv.org/...2501.12599
11. ^1 ^2 ^3 ^4"China's Moonshot AI Releases Trillion Parameter Model Kimi K2." HPCwire / AIwire, July 2025. hpcwire.com/...es-trillion-parameter-model-kimi-k2
12. ^1 ^2 ^3 ^4"Moonshot Releases Kimi K2, a Trillion-Parameter Model Fine-Tuned for Agentic Tool Use." The Batch, DeepLearning.AI. deeplearning.ai/...fine-tuned-for-agentic-tool-use
13. ^1 ^2 ^3 ^4"Moonshot Founder Reveals Kimi's Technical Roadmap: Token Efficiency, Long Contexts, and Agent Swarms." TMTPost, 18 March 2026. en.tmtpost.com/...7919010
14. ^1 ^2"Moonshot AI Secures $300 Million in Latest Funding Round, Boosting Valuation to $3.3 Billion." Pandaily. pandaily.com/...-boosting-valuation-to-3-3-billion
15. ^1 ^2 ^3"China AI Startup Moonshot Seeks $10 Billion Value in New Funding." Bloomberg, 17 February 2026. bloomberg.com/...s-10-billion-value-in-new-funding
16. ^1 ^2 ^3"Scaling law remains valid, but priorities are shifting, says Moonshot AI founder." KrASIA, November 2024. kr-asia.com/...e-shifting-says-moonshot-ai-founder
17. ^1 ^2 ^3 ^4 ^5 ^6"China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets." TechCrunch, 7 May 2026. techcrunch.com/...nd-for-open-source-ai-skyrockets
18. ^"China's Moonshot releases a new open source model Kimi K2.5 and a coding agent." TechCrunch, 27 January 2026. techcrunch.com/...del-kimi-k2-5-and-a-coding-agent
19. ^"Moonshot AI Releases Kimi K2.5 - 1T Parameter Native Multimodal Agent Model." ComfyUI Wiki, 27 January 2026. comfyui-wiki.com/...-moonshot-ai-kimi-k2-5-release
20. ^Moonshot AI. "Kimi-K2.6" model card, Hugging Face. huggingface.co/...Kimi-K2.6
21. ^1 ^2 ^3"Moonshot AI's AGI Vision" (translation of Overseas Unicorn's 21 February 2024 interview with Yang Zhilin). ChinaTalk, 19 March 2025. chinatalk.media/...moonshot-ais-agi-vision
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Zhilin Yang. Personal homepage. kimiyoung.github.io (accessed 23 September 2026)
23. ^1 ^2 ^3 ^4Yang, Zhilin. "Advances in Generative Feature Learning." PhD thesis, CMU-LTI-19-010, Language Technologies Institute, Carnegie Mellon University, 2019. lti.cs.cmu.edu/...yang-zhilin-thesis.pdf
24. ^1 ^2Dai, Zihang; Yang, Zhilin; Yang, Yiming; Carbonell, Jaime; Le, Quoc V.; Salakhutdinov, Ruslan. "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context." arXiv:1901.02860, ACL 2019. arxiv.org/...1901.02860
25. ^Du, Zhengxiao; Qian, Yujie; Liu, Xiao; Ding, Ming; Qiu, Jiezhong; Yang, Zhilin; Tang, Jie. "GLM: General Language Model Pretraining with Autoregressive Blank Infilling." arXiv:2103.10360, ACL 2022. arxiv.org/...2103.10360
26. ^Jiang, Ben. "Moonshot AI founder's dispute with 5 investors up for arbitration in Hong Kong." South China Morning Post, 10 December 2024. scmp.com/...pute-5-investors-arbitration-hong-kong
27. ^1 ^2 ^3 ^4 ^5Lunden, Ingrid. "China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context." TechCrunch, 21 February 2024. techcrunch.com/...moonshot-ai-funding-china
28. ^1 ^2"China's Moonshot AI raises US$500 million in latest funding round: report." South China Morning Post, 1 January 2026. scmp.com/...00-million-latest-funding-round-report
29. ^1 ^2"Moonshot AI targets US$12 billion valuation as overseas revenue surges for Kimi models." South China Morning Post, 18 February 2026. scmp.com/...on-overseas-revenue-surges-kimi-models
30. ^Moonshot AI. "Kimi K3 Tech Blog: Open Frontier Intelligence." 16 July 2026. kimi.com/...kimi-k3
31. ^1 ^2 ^3"月之暗面完成超35亿美元F轮融资：估值超350亿美元，G轮或下周关闭" (Moonshot AI completes F round of more than US$3.5 billion). The Paper, 29 July 2026. thepaper.cn/newsDetail_forward_33682828
32. ^1 ^2"月之暗面Kim完成股改：创始人杨植麟担任董事长兼经理" (Moonshot AI completes joint stock conversion; founder Yang Zhilin becomes chairman and manager). Sina Technology, republishing MyDrivers, 30 July 2026, citing Tianyancha registration data. finance.sina.com.cn/...doc-inikqqhm1884301.shtml
33. ^1 ^2Moonshot AI. "Kimi-K2-Instruct" model card, Hugging Face. huggingface.co/...Kimi-K2-Instruct
34. ^Moonshot AI. "Kimi-K2-Thinking" model repository (created 4 November 2025), Hugging Face. huggingface.co/...Kimi-K2-Thinking
35. ^Moonshot AI. "Kimi-K2.7-Code" model repository (created 11 June 2026), Hugging Face. huggingface.co/...Kimi-K2.7-Code
36. ^1 ^2Moonshot AI. Kimi K3 repository and model card, GitHub (github.com/...Kimi-K3) and Hugging Face (huggingface.co/...Kimi-K3). Accessed 23 September 2026.