Yang Zhilin's Moonshot goes open-weight with Kimi K2 after a $2B round

Late in 2024, a Chinese startup that had built its name on reading very long documents suddenly changed course and started giving its models away. Yang Zhilin (杨植麟), the Tsinghua-trained researcher who founded Moonshot AI, was betting that openness, not secrecy, would win the next phase of the large-model (大模型) race.

The bet has since been backed by serious money. In May 2026, Moonshot closed a US$2 billion round led by Meituan's Long-Z Investments at a US$20 billion valuation — reported as the largest single raise in China's large-model sector. Yang, still in his early thirties, has become one of the most watched entrepreneur (创业者) figures in Chinese AI.

Key takeaways

• Moonshot AI was founded in Beijing in March 2023 by Yang Zhilin with co-founders Zhou Xinyu and Wu Yuxin, all Tsinghua University alumni.
• Its Kimi chatbot launched in October 2023 with a 200,000-Chinese-character context window, then expanded to 2 million characters in March 2024.
• Kimi K2 (July 2025) is a roughly 1-trillion-parameter mixture-of-experts model with about 32 billion active parameters, released open-weight under a modified MIT license.
• A May 2026 round of US$2 billion at a US$20 billion valuation was led by Meituan's Long-Z Investments.
• Yang Zhilin is a former Tsinghua professor with a Carnegie Mellon PhD who worked at Google Brain and Meta AI.

A researcher who picked a narrow wedge

Yang's background reads like a research résumé before it reads like a founder's. He earned a PhD at Carnegie Mellon University, worked at Google Brain and Meta AI, and returned to China to become a professor at Tsinghua University before starting Moonshot. That research lineage shaped the company's first product decision.

Rather than compete head-on with general chatbots, Moonshot chose a specific wedge: context length. When Kimi Chat launched in October 2023, it could ingest around 200,000 Chinese characters in a single session — among the longest context windows available in a consumer chatbot at the time. For users drowning in PDFs, contracts and papers, that single feature was the hook. By March 2024, Moonshot pushed the window to roughly 2 million characters.

The pivot to open weights

For its first two years, Kimi ran as a closed product; training details stayed secret. That changed after DeepSeek demonstrated that a comparatively low-cost open system could capture global attention. In response, Moonshot began publishing its models.

The decisive move came in July 2025 with Kimi K2, a mixture-of-experts model of roughly one trillion parameters (about 32 billion active at inference) released under a modified MIT license. K2 was positioned for demanding work — software development and financial research — and Moonshot followed it with reasoning and multimodal variants. In November 2025, the company reported that its K2 Thinking variant scored 44.9% on Humanity's Last Exam, 60.2% on BrowseComp and 71.3% on SWE-Bench Verified, figures it cited to argue that open-weight models could match closed frontier systems on agentic tasks.

Why a closed product opened up

Yang's logic was pragmatic. A startup cannot out-spend the hyperscalers on closed-model marketing, but it can win developers. By open-sourcing weights, Moonshot turned Kimi from a consumer app into a platform that third parties could embed, fine-tune and ship. The strategy also fed distribution: open models travel through developer channels that paid ads cannot buy.

The approach has shown commercial traction. Moonshot has been reported to surpass US$200 million in annualized recurring revenue by early 2026, as API usage grew alongside the K2 family. The company has also been reported to be preparing a Hong Kong listing, though that plan has not been formally confirmed.

The funding arc

Moonshot's capital story mirrors the sector's swings:

• February 2024 — a US$1 billion round led by Alibaba at a US$2.5 billion valuation.
• August 2024 — a US$300 million round at a US$3.3 billion valuation, with Tencent and Gaorong Capital among investors.
• May 2026 — a US$2 billion round led by Meituan's Long-Z Investments at a US$20 billion valuation.

The jump from a US$3.3 billion mark in 2024 to US$20 billion two years later reflects both Kimi's user growth and a broader re-rating of Chinese large-model assets after the open-weight wave.

Honest limitations

• The benchmark scores (Humanity's Last Exam 44.9%, BrowseComp 60.2%, SWE-Bench Verified 71.3%) are company-reported figures for K2 Thinking; independent reproduction should be checked before treating them as settled.
• The US$20 billion valuation and US$2 billion raise are based on reported financing; exact terms and investor lists are not fully public.
• The "Hong Kong IPO preparation" is reported but not officially filed or confirmed.
• Annualized recurring revenue above US$200 million is a reported estimate, not an audited disclosure.
• Yang's academic titles and prior roles (Tsinghua professor, CMU PhD, Google Brain, Meta AI) are well documented, but his exact tenure lengths vary across sources.

What readers can do now

1. If you build with LLMs, test Kimi K2's open weights directly — the modified MIT license lets you self-host for many commercial uses without a per-call fee.
2. If you follow the sector, watch whether Moonshot's open-weight strategy converts developers into durable revenue; that is the real test of Yang's pivot.
3. If you research long-context AI, use Kimi's 2-million-character window as a reference point for what production retrieval systems can already handle today.