Glossary
What is Kimi AI
In short
Kimi AI is an AI assistant and a family of language models from the company Moonshot AI, and some of the models are released with open weights.
Models and products · Kimi
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a huge school where every student has a favorite subject. A history question comes in, and the history expert answers; a math problem comes in, and the math expert takes it, while the others rest. So the school knows a great deal and does not get tired. Some smart programs work in a similar way: only the needed part works on each question.
In plain words
Kimi AI is a helper from Moonshot AI: people chat with it on the website, and developers connect its models through an API. Kimi models reason step by step, write code and carry out tasks as an agent: they look for information, call tools and assemble the result. Many versions are open, so you can download them and run them yourself. For example, a development team installs a model on its own servers and asks it to review bugs in the code without sending the source code outside.
For business
A company has two paths. The cloud and the API work under a contract with Moonshot AI PTE LTD in Singapore: there is no separate list of countries, but the terms forbid use of the service in countries under comprehensive sanctions and by persons under sanctions restrictions (checked 16 October 2026). A lawyer decides how this applies to your company. The second path is open weights. Kimi K2 is released under a modified MIT license and K3 under its own license, and its extra conditions do not apply to internal use. The larger models are huge, so first calculate what a million tokens will cost on your own hardware.
For a specialist
Kimi K2 (Kimi Team, 2025) is a Mixture-of-Experts (MoE) model: 1 trillion parameters, 32 billion of them active per token, 384 experts plus one shared expert, and MLA (multi-head latent attention). It was pretrained on 15.5 trillion tokens with the MuonClip optimizer, an extension of Muon with the QK-clip technique against training instability. Then it was fine-tuned on synthetic agent data and with reinforcement learning in real and synthetic environments. According to its model card, K3 has 2.8 trillion parameters with 104 billion active, picks 16 of 896 experts, alternates KDA (Kimi Delta Attention) linear attention layers with MLA layers, understands images and holds a context of about a million tokens. The recommended engines are vLLM and SGLang. The K3 license requires naming the model in the interface of a product with more than 100 million users a month or revenue above 20 million US dollars a month.
Example
A software studio runs an open model of this family on a server in a Russian data center and connects it to a code editor. The clients’ source code stays inside the closed network, and the team checks quality on its real tasks.
Related terms
Questions and answers
Sources
- Kimi Team, Kimi K2: Open Agentic Intelligence, 2025arxiv.org
- Hugging Face, model card moonshotai/Kimi-K2-Instruct, Modified MIT Licensehuggingface.co
- Hugging Face, model card moonshotai/kimi-k3, checked 16 October 2026huggingface.co
- Moonshot AI, Kimi K3 Licensehuggingface.co
- Moonshot AI, Kimi OpenPlatform Terms, version of 30 July 2026, checked 16 October 2026platform.kimi.ai
Not sure whether your company needs this?
AI solutions map