Yang Zhilin published two of the papers that shaped how modern language models handle long sequences before he turned thirty, then left research to build the company that would ship them at scale. He is the co-founder and chief executive of Moonshot AI, the Beijing lab behind the Kimi model family, and in July 2026 he oversaw the release of Kimi K3, the first openly downloadable model to place inside the global top three on independent intelligence rankings.
He bylines his research as Zhilin Yang and goes by Yang Zhilin in Chinese order. Born in 1992 in Shantou, Guangdong, he was 34 when K3 shipped.
Tsinghua, a rock band, and Carnegie Mellon in four years
Yang entered Tsinghua University in thermal engineering and switched to computer science after reading Haruki Murakami, an origin story he has repeated in Chinese press. He graduated in 2015 under adviser Jie Tang, who later co-founded the rival lab Z.ai. He also drummed and wrote for a campus rock band called Splay, named after the splay tree data structure. Two of his bandmates would later co-found Moonshot with him.
He completed a PhD at Carnegie Mellon's Language Technologies Institute in four years, finishing in 2019 under Ruslan Salakhutdinov and William W. Cohen. Carnegie Mellon confirmed the 2019 date to The National when K3 made him newsworthy outside China. His academic homepage is still online at kimiyoung.github.io, last updated around 2022.
The research that made his name
Yang is a first author on Transformer-XL, which introduced segment-level recurrence and relative positional encoding so transformers could handle sequences longer than a fixed window. He is also a first author on XLNet, the permutation language modeling approach that beat BERT across twenty benchmarks in 2019. His Google Scholar record shows citations in the tens of thousands.
The through line between that work and Moonshot's product strategy is direct. Kimi launched in 2023 on a long-context pitch, and K3 ships with a one-million-token window. Yang built the company around the problem he had spent his PhD on. He interned at Google Brain and worked with Meta's FAIR during that period, co-authoring with Yoshua Bengio and Yann LeCun, and he remains a listed author on Moonshot papers including the widely cited Muon optimizer work.
When K3 landed, his former adviser Salakhutdinov posted publicly that Yang had made fundamental contributions to machine learning during his time at CMU, a quote picked up by Fortune.
Two companies, one thesis
Before Moonshot he co-founded Recurrent AI in 2016, a sales technology company built on Transformer-XL that applied language models to enterprise conversation data. A dispute with five Recurrent AI investors, including GSR Ventures, went to arbitration in Hong Kong in December 2024, reported by the South China Morning Post. He also worked on an early version of Huawei's PanGu model in 2020 and led a Wu Dao team at BAAI in 2021.
ChatGPT changed his plan. He traveled to the United States in December 2022 to study what OpenAI had shipped and founded Moonshot the following March. The stated goal is artificial general intelligence via three milestones: long context, a multimodal world model, and an architecture that can keep improving without human supervision.
The openness bet
Yang's competitive position is that giving the weights away buys distribution faster than a closed API sells it. Bloomberg reported he wants to grow the user base through openness and broader availability than the American proprietary systems allow, a framing echoed in Quartz. Moonshot still charges for hosted inference and consumer subscriptions, so the weights are the funnel rather than the product.
He was the only independent Chinese LLM founder invited to speak in person at Nvidia GTC in March 2026, where his talk covered scaling Kimi K2.5, the Muon optimizer, and linear attention for long-running agents. For builders deciding what to run, our guide to the best local LLMs covers where his models fit against the alternatives, and our report on the K3 release covers what it changed.
Frequently asked questions
What is Yang Zhilin known for?
Two things. Academically, he is a first author on Transformer-XL and XLNet, both foundational papers on long-sequence language modeling. Commercially, he is the co-founder and CEO of Moonshot AI, the Beijing lab that builds the Kimi models.
Where did Yang Zhilin study?
He studied computer science at Tsinghua University, graduating in 2015, then completed a PhD at Carnegie Mellon University's Language Technologies Institute in 2019 under Ruslan Salakhutdinov and William W. Cohen. He finished the doctorate in four years.
Does Yang Zhilin have a public social media account?
No personal account has been verified. Moonshot posts through the company account at @Kimi_Moonshot, and his academic work is listed on his own site and his GitHub profile. Treat accounts claiming to be him personally with caution.
What company did Yang Zhilin run before Moonshot AI?
Recurrent AI, founded in 2016, which applied his Transformer-XL research to enterprise sales conversation data. He also contributed to Huawei's early PanGu model and led a Wu Dao team at the Beijing Academy of Artificial Intelligence.
Why is Moonshot AI named after a Pink Floyd album?
The Chinese name 月之暗面 means the dark side of the moon. Yang has named the album as a favorite, and the company launched in March 2023 on its fiftieth anniversary. The band connection is not incidental: he played in a rock band at Tsinghua with two of his co-founders.