Moonshot AI is the Beijing lab behind the Kimi family of models, and as of July 2026 it occupies a position no Chinese startup held two years earlier: an openly downloadable model sitting inside the global top three. Kimi K3, released on July 27, 2026, is a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and a one-million-token context window. Anyone can download the weights. The American labs it competes with sell API access instead.
The company was founded in March 2023 by Yang Zhilin and two Tsinghua University classmates, Zhou Xinyu and Wu Yuxin. Its Chinese name, 月之暗面, translates as the dark side of the moon, and the launch was timed to the fiftieth anniversary of the Pink Floyd album Yang has named as his favorite. Three years on it carries a $20 billion valuation and annual recurring revenue that passed $200 million in April 2026, per TechCrunch, and reportedly $300 million by June.
From a long-context chatbot to a frontier lab
Moonshot shipped its first product, the Kimi chatbot, in October 2023. The pitch was context length: 200,000 Chinese characters per conversation at launch, raised to two million by March 2024. That was a narrow wedge, and it worked. Kimi became one of the most used consumer AI assistants in China before the company had a reasoning model to its name.
The research followed. Kimi K1.5 landed in January 2025 claiming parity with OpenAI o1 on math and coding. Kimi K2 arrived in July 2025 as a one-trillion-parameter mixture-of-experts model with 32 billion active parameters, published under a modified MIT license, and drew coverage in Nature. Kimi K2 Thinking followed in November 2025, adding 256,000 tokens of context, native INT4 inference, and the ability to chain two to three hundred sequential tool calls. Moonshot said it trained for about $4.6 million.
Kimi K2.5 brought native vision in January 2026 through a 400-million-parameter MoonViT encoder. K2.6 and the code-focused K2.7 followed through the spring and summer. By May 2026 Kimi was the second most used model on OpenRouter.
What Kimi K3 actually ships
K3 is not an incremental bump. Its model card describes 896 experts with 16 selected per token, a new attention design the lab calls Kimi Delta Attention paired with attention residuals, and native MXFP4 weights trained with quantization awareness from the supervised fine-tuning stage forward. Moonshot claims roughly 2.5 times the scaling efficiency of K2.
The reported numbers put it in frontier company: 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp. Independent trackers placed K3 third on the Artificial Analysis intelligence index and first in the Frontend Code Arena. Writing at Interconnects, Nathan Lambert argued the lag between open weights and closed frontier models has compressed from six to nine months down to three to five.
One caveat worth reading carefully: K3 does not ship under MIT. It carries a custom Kimi K3 License, a step back from the permissive terms K2 used. Check the license text before you build a commercial product on it.
Who funds Moonshot, and why Alibaba is both backer and rival
Moonshot raised about $2 billion at a $20 billion valuation in May 2026, in a round led by Meituan venture arm Long-Z with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. The valuation ladder before that ran $2.5 billion in early 2024, $3.3 billion later that year, $10 billion in early 2026, and $18 billion by March. Alibaba, Tencent, HongShan, IDG Capital, and ZhenFund are all on the cap table.
Alibaba holds roughly 36 percent of the company and also competes with it directly through the Qwen model family. Days after K3 shipped, Alibaba announced a 2.4-trillion-parameter open-weight Qwen model of its own. That arrangement is unusual by Silicon Valley norms and normal by Chinese ones, where the large platforms fund the labs that will eventually take their customers.
Why open weights, and what it means for builders
Moonshot gives away the thing that costs the most to make. The reasoning, per Yang, is distribution: openness buys a user base and a developer ecosystem faster than a closed API can. Our guide to the best local LLMs you can run at home covers where K3 fits for people running models on their own hardware, which is honestly nowhere, since 2.8 trillion parameters will not load on a desk. The practical path is renting GPU time by the hour or running a smaller distill locally.
The strategic effect is larger than any single download. Our news coverage of the K3 release laid out the pressure it creates: if a downloadable model performs within a few points of the best closed model, the pricing power of a frontier API narrows to the gap. Moonshot charges for hosted inference and consumer subscriptions and paused new Kimi signups after K3 launched because demand outran capacity.
The friction worth knowing about
In February 2026 Anthropic accused Moonshot, alongside DeepSeek and MiniMax, of using fraudulent accounts to distill Claude outputs, a claim reported by The New York Times. Beijing called the allegations groundless. No finding has been published either way, and the dispute has not slowed adoption.
The counterweight is that serious American products build on Moonshot models anyway. Cursor shipped on K2.5. Moonshot also publishes real research rather than only model cards: its Mooncake serving paper won a best paper award at USENIX FAST, and the Muon optimizer paper it co-authored with UCLA is widely cited. This is a lab, not a repackager.
Frequently asked questions
Who founded Moonshot AI?
Yang Zhilin, Zhou Xinyu, and Wu Yuxin founded the company in Beijing in March 2023. All three studied at Tsinghua University, and Yang serves as chief executive. Yang holds a PhD from Carnegie Mellon and co-authored the Transformer-XL and XLNet papers before starting the company.
Is Kimi K3 free to use?
The weights are free to download from Hugging Face, but the license is a custom Kimi K3 License rather than MIT, so commercial terms need reading. Hosted API access through Moonshot is paid, and consumer Kimi subscriptions are sold in tiers in China.
Can I run Kimi K3 on my own computer?
Not on consumer hardware. At 2.8 trillion total parameters the full model needs datacenter-class memory even at 4-bit quantization. Most people either call the hosted API, rent GPU time by the hour, or run a smaller open model locally instead.
How does Moonshot make money?
Two revenue lines. Consumer subscriptions to the Kimi assistant, sold in tiers in the Chinese market, and paid API access for developers through its platform. TechCrunch reported annual recurring revenue passed $200 million in April 2026.
How does Moonshot compare to DeepSeek?
Both are Chinese labs publishing open weights, and both have had moments where a release reset expectations about the open-to-closed gap. DeepSeek has historically carried the higher public profile outside China. K3 in July 2026 was widely described as Moonshot's turn at that moment.