1. Yang Zhilin, founder of Moonshot AI, has become a prominent figure in China's new generation of AI entrepreneurs, combining elite academic credentials with an unconventional career path. Moonshot AI released Kimi K3 in July—an open-weight model with 2.8 trillion parameters that set a record for model size—while the company rapidly raised capital and prepared for a Hong Kong listing. Corporate filings show Yang, now 33, owns about 51.8% of Moonshot AI's domestic entity as its largest shareholder. [para. 1][para. 2][para. 3]
2. Born in Shantou, Guangdong, in November 1992, Yang nearly pursued humanities before being selected for his high school's informatics-olympiad program, which led to Tsinghua University. He entered Tsinghua in 2011, initially studying thermodynamics, but was inspired to pursue coding after reading about a programmer in a Haruki Murakami novel—fascinated by "writing code late at night and turning technology into something real." He switched to computer science and joined the lab of Tang Jie, who later co-founded Z.AI. Tang called Yang "the best and most talented student" he had seen in years, and Yang consistently ranked first in his class with full marks in every programming course. [para. 4][para. 5][para. 7][para. 8][para. 9]
3. Yang also formed a rock band at Tsinghua called Splay—named after the splay-tree data structure—serving as drummer and songwriter; their song "Bleeding" won Tsinghua's best-original-song award in 2014. Music occasionally crossed into research: he once solved a stubborn mathematical derivation while walking back to his dorm listening to a favorite song. Two Tsinghua relationships became central to Moonshot AI: Splay guitarist Zhou Xinyu became co-founder for model training, and classmate Zhang Yutao became co-founder and CTO. [para. 10][para. 11][para. 12]
4. After graduating first in his class, Yang earned a doctorate at Carnegie Mellon under advisers Ruslan Salakhutdinov (then head of Apple AI research) and William Cohen (principal scientist at Google), interning at Meta AI Research and Google Brain. He has published over 20 papers with more than 52,000 citations; his best-known works Transformer-XL and XLNet (2019) influenced natural-language processing, and he also published with Turing Award winners Yann LeCun and Yoshua Bengio. He completed his doctorate in four years, and Salakhutdinov described his work as making foundational contributions to machine learning. [para. 13][para. 14][para. 15][para. 16]
5. Despite Apple's attempts to recruit him—including a position in its China operation—Yang chose entrepreneurship, believing large tech companies were too slow to turn AI research into products. His first startup, Recurrent AI (with Zhang Yutao and Carnegie Mellon alumnus Chen Qicong), raised nearly $100 million from investors including HSG, ZhenFund, GSR Ventures and Boyu Capital. Recurrent AI also worked with Huawei Cloud on the Pangu NLP model (2021, billed as the industry's first 100-billion-parameter Chinese-language model) and Yang participated in the Beijing Academy of AI's Wu Dao project. [para. 17][para. 18][para. 19][para. 20]
6. After ChatGPT's launch, Yang saw late 2022 as a historical turning point—where high-quality internet data, mature computing infrastructure and effective transformer training had finally converged. In early 2023, Yang and Zhang left Recurrent AI to create Moonshot AI (named after Pink Floyd's "The Dark Side of the Moon"), with Recurrent AI receiving a stake. The separation led to a legal dispute: five Recurrent AI investors, including GSR Ventures, 5Y Capital and Boyu Capital, filed Hong Kong arbitration alleging Yang and Zhang began fundraising for Moonshot before obtaining required waivers; negotiations over equity continue. [para. 21][para. 22][para. 23]
7. Moonshot AI's early strategy centered on "long context" as a new form of computer memory. Kimi Chat launched in October 2023, initially handling 200,000 Chinese characters, then expanding to 2 million in early 2024—propelling Kimi into the mainstream. By November 2024, Kimi claimed 36 million monthly active users, and Moonshot raised over $1.3 billion in 2024 from Tencent, Alibaba, Monolith and Xiaohongshu. However, DeepSeek's advances in early 2025 shifted investor attention from user growth back to frontier-model capability, forcing Moonshot to cut advertising spending, abandon its K1 model and concentrate resources on a more ambitious K2. [para. 24][para. 25][para. 26][para. 27]
8. Kimi K2 arrived in July 2025 with 1 trillion parameters, then the largest open-weight model; Kimi K3 followed a year later with 2.8 trillion parameters. Yang has worked to make Moonshot's culture reflect interests beyond benchmark scores—before K3's release, the company posted a 36-second AI-generated video that online commenters praised, calling Moonshot "the most artistic" of China's large-model companies. Yang himself keeps a low public profile, rarely giving media interviews and maintaining no public social-media account. [para. 28][para. 29][para. 30]
9. Moonshot AI completed a $500 million Series C at the end of 2025; during 2026 its valuation climbed from about $8 billion to a $50 billion pre-money valuation within roughly six months as it prepared for a Hong Kong IPO. The trajectory gives ironic resonance to lyrics Yang wrote in 2014 for a Splay song called "Daydream"—imagining his company going public, counting money and buying a Lamborghini—a daydream that more than a decade later looks far less fanciful. [para. 31][para. 32][para. 33]
AI generated, for reference only
Sign in with Google
Sign in with Facebook
Sign in with 财新