From Changing Majors to Long-Sequence Research
Zhilin Yang went from Shantou to Tsinghua, initially studying thermal engineering before switching to computer science in his sophomore year. After graduating in 2015, he went to CMU for a PhD in language and information technology, studying generative feature learning and long-range dependencies under Ruslan Salakhutdinov and William W. Cohen, and participating in work such as Transformer-XL and XLNet.
This experience is not decorative biography. Kimi later entered the market through long context, backed by a traceable research continuity: first solve how models process longer dependencies, then rewrite that capability as “users can hand AI more material at once.”
The First Startup: Learning Industrialization
During his PhD, Zhilin Yang and his team began incubating Recurrent AI. The company later said that 2016–2017 was mainly spent exploring application directions and that formal business began in 2018. We should neither write 2016 as the birth of a mature company nor ignore its significance as a first entrepreneurial exploration.
Recurrent AI gave him the real constraints of enterprise AI: beyond model quality, there were data, delivery, sales processes, and customer ROI. In 2021 the company completed a $38 million financing, a clear institutional result, but he was a co-founder and technical core rather than a chief executive with every power concentrated in one person.
Becoming the Person in Charge at 30
After founding Moonshot AI in March 2023, Yang's identity changed in a decisive way. He was no longer merely bringing technology into a company; he had to decide how models, products, financing, and organization should be combined.
Kimi, released in October of the same year, did not vaguely copy a “Chinese ChatGPT.” Instead, it established a clear label with a context window of about 200,000 Chinese characters. The roughly $1 billion financing in 2024 validated the capital market's bet, but financing is not operating revenue. A dispute involving some early investors in Recurrent AI later entered arbitration; public materials show no final ruling, so this entry records only that a dispute existed and does not assign responsibility.
From Chinese Consumers to Global Developers
Kimi K2's open weights in 2025 changed the distribution logic again: from serving Chinese consumers directly to allowing global developers to continue building on the model. K3 in 2026 continued this route, after which Bloomberg reported that the company completed financing of about $3.5 billion at a post-money valuation of about $35 billion.
The figure is large, but the evidence boundary is equally clear: it comes from reporting by people familiar with the matter, not complete public financing documents, and certainly not Zhilin Yang's personal liquid wealth. What matters is whether he can turn the advantage of papers, the consumer entry point, and the open-model ecosystem into a sustainable operating system.