What is Moonshot AI, and how did Kimi get from K2 to K3?

Moonshot AI (in Chinese, Yuè Zhī Ànmiàn, "the dark side of the moon") was started in Beijing in March 2023. Its three founders, Yang Zhilin, Wu Yuxin and Zhou Xinyu, knew each other from Tsinghua University. It launched its Kimi chatbot in October 2023, and the name comes from Yang's English nickname. In early 2024, Alibaba led a funding round of about $1 billion US at a $2.5 billion valuation, according to reports summarized by Wikipedia.

Most of the company's outside reputation comes from its open-weight models. Open weight means the trained numbers of the model are published for download. Moonshot's repeated bet has been size: for nine of the twelve months before K3, it says, a Kimi model was the largest open model available.

Every major Kimi model so far

Kimi K1.5, announced in January 2025, came with a technical report describing its reinforcement learning methods, but no K1.5 repository appears on Moonshot's Hugging Face page. The open line for its flagship models starts with K2.

Model	When	Total / active parameters	Main addition	Licence
Kimi K2	July 2025	1 trillion / 32 billion	Pre-trained on 15.5 trillion tokens; built for agents and coding	Modified MIT
K2 update (0905)	September 2025	About the same	Revised K2 weights	Modified MIT
K2 Thinking	November 2025	About the same	Step-by-step reasoning between tool calls, 200 to 300 in a row; 256,000-token context	Modified MIT
K2.5	January 2026	About the same	Native vision, from about 15 trillion more mixed tokens; agent swarm of up to 100 sub-agents	Modified MIT
K2.6	April 2026	About the same	Longer coding runs; swarm of up to 300 sub-agents	Modified MIT
K2.7 Code	June 2026	About the same	Coding focus; about 30 percent less thinking output than K2.6	Modified MIT
K3	Launch July 16, weights July 27, 2026	2.8 trillion / 104 billion	New attention designs; 1 million tokens of context	Kimi K3 License
K3.1 Pro	October 2026	3T class	More post-training on K3; weights promised by November 4	Not yet published

A parameter is one learned number. In a mixture-of-experts model, only some of the parameters work on each token, a token being a small piece of text. That explains the two figures: active parameters drive the compute per token, and total parameters drive the memory you need.

How the licence changed

The K2 family uses a short "Modified MIT" licence. It is the standard MIT permission plus one sentence: a commercial product with more than 100 million monthly users, or more than $20 million US in monthly revenue, must prominently show the model's name.

The Kimi K3 License keeps that display rule and adds a second one. A business that sells Model as a Service, meaning API access where customers control inputs, parameters or training data, needs a separate agreement once its group's revenue passes $20 million US over any 12 months. Internal use and access through Moonshot's own products or certified partners are exempt.

Be careful with summaries of this licence. Some online pages say K3 is noncommercial, and others say it requires a revenue share of up to 30 percent. The licence file itself says neither: it calls for a separate agreement and states no percentage. Whatever terms that agreement holds are private, so read the file and, for a hosting business, ask Moonshot.

How the hardware grew

The size of the download tells the story, and these are the file totals on Hugging Face:

Model	Files	Why
Kimi K2 Instruct	About 1,029 GB	One trillion parameters at 8 bits
Kimi K2.6	About 595 GB	Similar size, with native INT4 quantization
Kimi K3	About 1,560 GB	2.8 trillion parameters in 4-bit MXFP4

Quantization means storing each number in fewer bits to save memory. It is why K2.6 needs less than K2 and why K3 is not far larger. Even so, none of these models fits a home computer at full quality. K3 sits firmly in data-centre territory, with vLLM's setup guide listing eight GB300 chips as a minimum.

What the debate around Kimi is about

K3 brought Moonshot into a political fight. The New Stack reported in July that Anthropic and US officials accuse the company of distilling outputs from American models, and that Moonshot denies it, crediting its new architecture. Wikipedia also summarizes a Bloomberg report on Moonshot's access to NVIDIA chips through Alibaba Cloud, which Alibaba denies, and pressure from a US House committee on American companies that use Kimi. These are allegations and reports, not findings. They matter if your organization has compliance rules about Chinese suppliers.

The weights themselves can be separated from the service. If you download K2.6 and run it on your own machines, your prompts do not go to Moonshot. If you use Kimi.com or Moonshot's API, the company's terms and data policies apply, so read them before sending anything confidential.

Try it: Pick one coding question you know the answer to and ask it on Kimi.com with K3, then on a hosted K2.6 through a provider you already use. Compare the answer, the time taken and the tokens billed.

Which Kimi model should you try today?

For a first look, use the Kimi site or app, since it needs no hardware. For a builder, K3 through the Kimi API uses OpenAI-style and Anthropic-style calls, according to its model card, so testing often means changing an address and a model name.

For self-hosting, K2.6 or K2.7 Code is the more realistic choice: a smaller download and the simpler Modified MIT licence, though you still need a server with several GPUs. If you want to run something on one machine at home, Kimi is not the family to start with; our open-weight model hub lists models built for that.

K3.1 Pro is the one to watch. Moonshot says its weights will be released by November 4, and until they appear, the licence and file sizes are unknown.