Aiodic

Back to the table

69 Model family Ki Kimi Model Families

Kimi

Moonshot AI's model family.

What it is

Kimi is the name of a chatbot and of the family of AI models behind it, both made by Moonshot AI, a start-up based in Beijing. Like GPT, Claude, and Gemini, the Kimi models are language models: systems that learned from huge amounts of text to predict and write words, and so can answer questions, summarise documents, write code, and carry out multi-step tasks.

Moonshot AI was founded in March 2023 by Yang Zhilin and two fellow graduates of Tsinghua University, Zhou Xinyu and Wu Yuxin. Yang had earned a PhD at Carnegie Mellon University, where he co-wrote two well-known research papers on language models, Transformer-XL and XLNet, and had worked at Google Brain and Meta. The company's name is a nod to Pink Floyd's album The Dark Side of the Moon, which came out 50 years before the company did.

Two things have defined Kimi from the start:

The generations, in brief

Model families change quickly, so treat this as a snapshot at the time of writing, with dates taken from Moonshot's own model pages and news reports:

Open weights, not fully open source

Moonshot describes its K-series models as "open source", but it is more precise to call them open weight. The weights are the billions of numbers a model learned during training. Moonshot publishes those, plus code to run them and technical reports, but not the full training data, so you can use and adapt the models but not rebuild them from scratch.

The licences also carry conditions that ordinary open-source licences don't:

Licence terms change between versions, so read the licence file of the exact model you download, and ask a lawyer if the stakes are high.

Being open weight doesn't mean you can run it at home. The original K2 download is about 1.1 terabytes. Unsloth, a company that makes compressed versions of open models, cut K2 Thinking down to about 245 GB, which still needs a machine with roughly that much memory and storage combined. For almost everyone, using Kimi means the Kimi app, the kimi.com website, or Moonshot's paid service for developers, known as an API. The Parameters page explains what model size means for running models on your own computer.

Why it matters

It showed that a start-up could release a model near the top of the field. When K2 came out in July 2025, the science journal Nature ran a news story headlined "'Another DeepSeek moment'", comparing it with the shock caused by DeepSeek's R1 earlier that year. That story noted that K2 did well on coding tests and cost much less to use through its API than leading closed models. Moonshot is a young company rather than a tech giant like Alibaba, maker of Qwen, which made the result more striking. The wider story of China's open-weight wave and US chip export controls is told on the DeepSeek and Qwen pages.

It is built into other people's products. Because the weights are downloadable and the licence is permissive for most users, other companies build on Kimi. Moonshot's own API also mimics the format Anthropic uses for Claude, and Moonshot publishes a guide to plugging Kimi models into Claude Code, Anthropic's coding tool, in place of Claude. That makes Kimi one of the cheaper engines people try inside coding agents. The Cursor example below shows how far this goes.

It pushed agent features into open models. Many agent features first appeared in closed models you can only rent. Kimi's step-by-step tool use, Agent Swarm, and its own command-line coding tool (released under the Apache 2.0 licence) put similar features into models anyone can download and inspect. Moonshot made these capability claims itself, so test them on your own work.

It raises real questions about trust. Moonshot is a Chinese company. Its models, and its apps in China, must follow Chinese content rules, and it has been at the centre of serious accusations from Anthropic, the maker of Claude, about copying. Anyone choosing a model should understand these issues as well as the strengths. The examples below cover both.

Real-world examples

The K2 release and the "DeepSeek moment" comparison (July 2025)

On 11 July 2025 Moonshot published Kimi K2 on Hugging Face, the main website where open models are shared, in two versions: a base model for developers to train further and an "Instruct" version tuned to follow requests. Its technical report said the model was trained on 15.5 trillion tokens using a new training method the team called MuonClip, and reported a score of 65.8% on SWE-bench Verified, a test of fixing real bugs in open-source software, in a single attempt. Those figures come from Moonshot, not an independent tester.

Within days, developers were trying it inside their tools. On 13 July 2025 the software developer Jesse Vincent wrote a blog post showing how to run Claude Code with K2 in place of Claude, using Moonshot's Claude-compatible API, and warned readers against unofficial setup scripts circulating online. Nature's news story followed later that month.

The independent view was more mixed. In December 2025 the US government's Center for AI Standards and Innovation (CAISI), part of the National Institute of Standards and Technology, evaluated K2 Thinking. It found the model was a modest step up on DeepSeek's V3.1 but still behind leading US models, even older ones. It also reported that a month after release, K2 Thinking had been downloaded far less than DeepSeek R1 or OpenAI's open model. Excitement among developers and actual adoption are not the same thing.

Cursor's Composer 2 turns out to be built on Kimi (March 2026)

In March 2026 Cursor, one of the most popular AI coding tools, launched its own coding model, Composer 2. Its announcement didn't say what the model was built on. Developers soon noticed references to Kimi in the model's identifiers, and Cursor confirmed that Composer 2 started from Kimi K2.5. As reported by eWeek, Cursor's Lee Robinson said only about a quarter of the computing power behind the final model came from the base model, with the rest from Cursor's own training, and co-founder Aman Sanger said it was "a miss" not to mention the Kimi base from the start. Moonshot's Kimi account said the use came through Fireworks AI as part of an authorised commercial partnership.

The episode shows how open-weight models work in practice: a company takes a downloaded model, trains it further for its own purposes, and sells the result under its own name. It also drew political attention. On 29 April 2026, two US House committee chairmen announced an inquiry into Anysphere, Cursor's parent company, and Airbnb over their use of Chinese models, noting that Composer 2 was reportedly built on a Moonshot model (the Airbnb side is covered on the Qwen page).

Anthropic's accusations of copying (2026)

On 23 February 2026, Anthropic said it had found "distillation attacks" on Claude by three Chinese labs: DeepSeek, Moonshot, and MiniMax. Distillation means training one model on the answers of another. Anthropic said Moonshot ran more than 3.4 million exchanges with Claude through hundreds of fraudulent accounts, focused on agents, tool use, coding, and computer vision, and that it linked the activity to Moonshot partly through request details that matched the public profiles of senior Moonshot staff.

On 10 September 2026 Anthropic went further. As reported by the South China Morning Post and others, its threat report alleged that Moonshot had quietly forwarded some Kimi users' requests to Claude, shown Claude's answers to those users as if they came from Kimi, and used the exchanges to train its own models: more than 23 million exchanges between May and July 2026, through 5,380 fraudulent accounts mostly appearing to be in Singapore and Japan. Decrypt reported that China's internet regulator, the Cyberspace Administration of China, then questioned Moonshot and DeepSeek over whether sensitive Chinese data had ended up on US servers as a result.

These are allegations by a direct competitor. At the time of writing they have not been tested in court, and we could not find a public response from Moonshot. According to Runtime Wire, China's commerce ministry rejected the broader US allegations as groundless. The point for readers isn't to settle the dispute here, but to notice that if the claims are true, a Kimi user may not have known which company's model was answering, or where their words went.

An everyday case: the tender document

Here is an illustrative scenario, not a real company. Ravi runs a small building firm and has a 300-page public tender document to respond to. He uploads it to the Kimi website and asks for a list of every deadline, every required certificate, and every penalty clause, each with its page number. Kimi returns a tidy list in under a minute. Ravi checks each item against the page it cites, finds that two page numbers are wrong and one clause was missed, and fixes them. The summary saves him an afternoon, but he doesn't upload his own pricing or staff details: he has read the privacy policy and decided company-sensitive material stays off any outside chatbot, whoever makes it.

Try it yourself

This exercise tests Kimi's best-known strength, reading long documents, and shows you how to check the answer. It uses a public document, so there's nothing private at stake. Menus change often, so the labels may differ a little.

  1. Pick a long public document you can check. A public-domain novel from Project Gutenberg (gutenberg.org) works well, for example Pride and Prejudice. Download the plain text or PDF version.
  2. Open kimi.com in your browser. At the time of writing it may ask you to sign in; free use has limits that change, so check the current terms. Anything you upload is processed on Moonshot's servers, so never upload anything private.
  3. Upload the file using the attachment button, and paste this prompt: > Using only the attached book, list every time a character receives a letter. For each one, give the chapter number, who sent it, and one sentence on why it matters to the plot. If you are not sure about any item, say so.
  4. Check three items yourself. Open the book, go to the chapters it named, and see whether each letter is really there. Long-context models can miss items or invent details, a problem the Hallucination page explains.
  5. Ask a follow-up that needs the whole book: > Which character changes their opinion of another character most over the whole book? Quote two short passages, one early and one late, that show the change, with chapter numbers.
  6. Run the same prompts in another chatbot you use, and compare. Which one found more letters? Which one quoted accurately?
  7. Look at the licence. Go to huggingface.co/moonshotai, open any K2 model, and click the LICENSE file. Find the sentence about 100 million monthly active users. This is what "open weight with conditions" looks like.
  8. Write down your rule. Finish this sentence: "I'd use a long-context chatbot for ___, but I'd always check ___ myself."

Common mistakes and misunderstandings

Thinking "open source" means no strings attached. Kimi's licences are permissive for most people, but they add display rules for very large products and, for K3, a separate agreement for big model-selling businesses. The training data isn't published either. Read the licence of the exact version you use.

Assuming open weight means you can run it on your laptop. Kimi's flagship models are among the largest open models ever released. Unless you have server-class hardware, you'll use them through Moonshot's app, its API, or a third-party provider. Smaller open families such as Qwen or Gemma are better for home use.

Mixing up the model with the app. When you download a Kimi model and run it yourself, or use it through a provider you trust, your data stays there. When you use kimi.com or the Kimi app, your data goes to Moonshot. At the time of writing, the international privacy policy names a Singapore company as the provider, says your information may be stored on servers outside your country, and says content is used to train and improve its models. The mainland China policy says data is stored in China. The Privacy page covers what to check.

Forgetting the political filter. Models made in China must follow rules on politically sensitive content (the DeepSeek page explains them). NIST's CAISI found that K2 Thinking was "highly censored in Chinese" but relatively uncensored in English, Spanish, and Arabic. For history, politics, or current affairs involving China, check other sources. Western models have their own biases too.

Treating long context as perfect memory. A model that can take in a whole book can still miss things in the middle of it. Ask for page or chapter references, and check them.

Taking benchmark claims at face value. Moonshot, like every model maker, publishes scores from tests it chose, and independent testers sometimes find a different picture, as CAISI did. Test on your own tasks, as the Evals and Benchmarks pages explain.

Assuming you always know which model answered. The Cursor episode showed that a product can run on Kimi without saying so, and Anthropic's allegations suggest the reverse can happen too. If it matters which model handles your data, ask the provider directly.

Expecting it to stay the same. Moonshot has released a major model every few months since 2025, with names, sizes, and licences changing each time. Anything here about versions or terms may be out of date by the time you read it.

Related elements