A chatbot that could read an entire book in one prompt sounded like a party trick in late 2023. Two years later, that same idea — long context — became the wedge a young Chinese researcher used to crack open the global large-model (大模型) race.
Most founders in this sector came from product or finance. Yang Zhilin came from a lab.
Key takeaways
- Yang Zhilin (杨植麟), born in 1992 in Shantou, founded Moonshot AI (月之暗面) in March 2023 after a research career spanning Tsinghua, Carnegie Mellon, Google Brain and Meta AI Research.
- The company's Kimi chatbot, launched in October 2023, stood out for handling roughly 200,000 Chinese characters in a single prompt.
- In July 2025 Moonshot released and open-sourced Kimi K2, a mixture-of-experts model with 1 trillion total parameters and 32 billion active per token.
- Alibaba led a US$1 billion round in February 2024 (valuation about US$2.5 billion); Tencent joined a US$300 million round in August 2024 (valuation about US$3.3 billion).
- Before Moonshot, Yang co-authored Transformer-XL and XLNet and co-founded Recurrent AI (循环智能) in 2016.
A researcher before a founder
Yang was not a typical startup founder. He studied at Tsinghua University and graduated in 2015 at the top of his computer-science class, then earned a PhD at Carnegie Mellon University in 2019 under Ruslan Salakhutdinov, a researcher who had also advised work behind modern deep learning. During his studies he became a first author on Transformer-XL and a co-author on XLNet — papers that, together with collaborators including Turing Award winners Yann LeCun and Yoshua Bengio, were cited tens of thousands of times and helped shape how language models handle long sequences.
He spent time inside two of the most influential AI labs in the world: Google Brain and Meta AI Research (FAIR). In 2016, before Moonshot, he co-founded Recurrent AI (循环智能), an enterprise speech-and-language company. That mix — academic depth plus a first venture — is unusual among the founders of China's large-model startups.
The long-context bet
When Moonshot launched in March 2023, the prevailing wisdom was that chatbots competed on chat quality. Yang pushed a different axis: length. Kimi, released in October 2023, could take in about 200,000 Chinese characters at once — enough for a short book, a long contract, or a year of chat logs. It was a deliberate choice to serve students, researchers and analysts who wanted to reason over dense documents rather than exchange small talk.
The product caught on. Users flooded Kimi for summarization and retrieval over long texts, and the company became one of a handful of Chinese challengers to Baidu's Ernie Bot and the overseas incumbents. But the rise of low-cost models from rivals in early 2025 pressured Moonshot's position, and the company responded by leaning back into research.
Kimi K2 and the open-source turn
In July 2025, Moonshot released Kimi K2 and published its weights. According to the company's own technical notes, K2 is a mixture-of-experts (MoE) model: 1 trillion total parameters, but only 32 billion active for each token, trained on 15.5 trillion tokens. The design trades raw size for efficiency — a smaller slice of the network fires per request, cutting cost while keeping capability. State-media technology reporting noted that Nature's website described the release as "another DeepSeek moment," and that K2 reached the top of open-model leaderboards on coding and agentic benchmarks.
That open-source move mattered beyond China. It placed Moonshot alongside DeepSeek, Alibaba and others releasing permissively licensed weights, a contrast with most U.S. frontier labs that keep their best models closed. For global developers, K2 became a way to build agents and coding tools on a Chinese-open model without asking permission.
Money and backers
Moonshot raised at a pace that matched its ambition. Alibaba led a roughly US$1 billion round in February 2024 at about US$2.5 billion valuation, with investors including Sequoia China (HongShan) and Xiaohongshu. In August 2024 Tencent joined a roughly US$300 million round that lifted the valuation to about US$3.3 billion. The roster of backers — Alibaba, Tencent, Meituan's investment arm and major venture funds — reflects how seriously China's internet incumbents took the large-model race.
Why it matters globally
Yang's path is a useful case study in how China's AI talent pool has matured. A founder with first-rate international training chose to build in Beijing, ship a consumer product, and then give away a trillion-parameter model. The result is a company that competes on both the application layer (Kimi) and the infrastructure layer (open weights) — a dual strategy few startups anywhere can sustain.
Honest limitations
This profile relies on company statements, state-media technology reports, and secondary profiles; it does not include independent benchmarking of K2's claims, which shift as the model updates. Funding figures reflect widely reported rounds through 2024 and may not capture later, less-documented raises. Birth year and early education are drawn from Chinese biographical entries and a company-backed profile; minor discrepancies exist across sources. I have excluded unverified legal disputes and market rumors.
What readers can do now
- Try Kimi K2's open weights or API for a document-heavy task — paste a long report and ask for structured analysis to see the long-context advantage directly.
- If you build agents, compare K2 against a closed model on a coding benchmark you care about rather than trusting headline scores.
- Track Moonshot's licensing terms before shipping a commercial product; open weights still carry usage conditions.