Kimi

Most assistants present a text box and let you discover what they can do. Kimi presents a row of outputs — slides, sheets, docs, websites, design — and asks which one you want.

That is a small interface decision carrying a large argument: that the useful unit of AI work is a finished artefact rather than a conversation you then copy out of.

Behind it sits Moonshot AI's K3 model, and the homepage says what it is for in the title: agentic coding and knowledge work.

Kimi: Moonshot AI's Agentic Assistant

• Built on K3, positioned for agentic coding and knowledge work rather than general conversation.
• Output surfaces as entry points: slides, sheets, docs, websites, design and deep research.
• Swarm and scheduled tasks — multiple agents, and work that runs when you are not there.
• Instant and High modes, letting you trade speed against depth per request.
• Operated by Moonshot AI with Chinese regulatory filings displayed, which belongs in any data decision.

What Does K3 Change For Knowledge Work?

The page title states the positioning directly: Kimi AI with K3, built for agentic coding and knowledge work.

That is a narrower claim than most assistants make, and the narrowing is the interesting part.

General assistants are optimised for the average request, which means being adequate at everything and shaped by nothing. Naming two domains — coding and knowledge work — implies choices about long-context handling, tool use and multi-step execution that a general model would balance differently.

Agentic is the load-bearing word. A conversational model answers; an agentic one takes an instruction, decomposes it, uses tools, and returns when the task is done. The interface reflects that, offering to task an agent alongside asking anything.

What the site does not provide is any benchmark, comparison or technical description of K3. The claim is positioning rather than evidence, and anyone choosing a model for coding should test it on their own repository rather than on a title tag — which is precisely the argument Warp makes about benchmarking on your own tasks.

A conversational model answers. An agentic one decomposes the instruction, uses tools, and comes back when the work is done.

Instant Versus High

Two modes sit beside the input: Instant and High. Neither is explained on the page, and the naming makes the trade obvious enough that it does not need to be.

This pattern has become standard across assistants for a good reason. Most requests are trivial — a definition, a rewrite, a quick lookup — and spending twenty seconds of reasoning on them is waste. A minority are genuinely hard and benefit from every second.

Exposing the choice to the user rather than deciding automatically is the right call, because only you know whether this particular question matters.

The failure mode is habitual selection. People pick one mode, forget the toggle exists, and either wait unnecessarily on simple things or get shallow answers to hard ones.

Worth building the habit deliberately: default to Instant, and switch to High when you notice yourself about to check the answer manually. That instinct is a reliable signal that the question deserved more.

Slides, Sheets, Docs And Swarm

The surfaces listed in the interface are Slides, Swarm, Deep Research, Docs, Websites, Sheets and Design, alongside Kimi Work, Kimi Code, Kimi Claw, Projects and Scheduled Tasks.

Reading that list as a product strategy tells you more than any feature description.

Five of those produce documents in formats people actually deliver — a deck, a spreadsheet, a document, a website, a design. The assistant is being positioned as something that hands you a file rather than text to reformat, and that is where most of the real friction in assistant use has always been.

Deep Research and Swarm are the agentic pair. Deep research means a long-running investigation returning a report; Swarm implies multiple agents working in parallel, which is the current frontier and the least proven.

Scheduled tasks are the quietest and most consequential. Work that runs on a schedule without you present changes an assistant from a tool you use into a process that runs — and that is a different thing to trust, monitor and be responsible for.

• Start from the output surface, not the chat box.
• Default to Instant and escalate to High deliberately.
• Treat scheduled tasks as automation, with the oversight that implies.
• Establish where your data goes before pasting anything sensitive.
• Test coding claims on your own repository, not on the marketing.

Where Kimi Is Operated From

The footer carries a Moonshot AI copyright alongside three Chinese regulatory registration numbers — an ICP filing, a B2 licence and a public security network registration.

This is a fact about the service rather than a judgement, and it is one that belongs in an adoption decision.

Those filings indicate a service operated under Chinese jurisdiction, which determines the legal framework governing data handling, retention and lawful access requests. That is the same class of consideration as choosing between a US and an EU-hosted vendor, and it is answered by different law.

For an individual asking general questions, the practical weight of that is low. For anyone pasting client material, unreleased product information, personal data covered by GDPR, or code under a confidentiality obligation, it is a governance question that needs answering before use rather than after.

The site does not publish a data processing agreement, retention policy or regional hosting option on the pages we read. Anyone with a compliance requirement should establish those directly, and anyone without one can reasonably proceed — the point is to make the decision consciously rather than by default.

An Interface That Is The Whole Homepage

There is no marketing site. Visiting kimi.com opens the product, with the model selector, the surfaces and an input box, and nothing selling you on any of it.

That is unusual and worth reading in both directions.

The generous interpretation is confidence: a product good enough that using it beats reading about it, and a company that would rather spend the first thirty seconds demonstrating than persuading. For an assistant, where the experience is the argument, that logic holds.

The cost is that a prospective user cannot answer basic questions without an account — what the pricing is, what data policies apply, what K3 actually is, what Swarm does, or what distinguishes Kimi Work from Kimi Code.

For anyone evaluating tools for an organisation rather than for themselves, that is a real obstacle. The information a procurement process needs does not exist on the site, and getting it means asking rather than reading — which is a choice the company has made and worth knowing before you start. A self-hosted alternative like LobeHub answers those questions differently, and Khoj differently again.

Putting An Agentic Assistant To Work

The way to waste an agentic assistant is to use it as a chat box, which is what almost everyone does out of habit.

This sequence gets at what the interface is actually built for.

1. Decide your data position before pasting anything work-related.
2. Pick the output surface first — slides, sheet, doc — rather than describing it.
3. Give a task with a finish condition, not a question.
4. Use High mode when you would otherwise verify the answer yourself.
5. Try one scheduled task and review what it produced unattended.
6. Test the coding claims against a repository you know well.

Kimi: Surfaces And Operator