The short answer
Kimi is the AI assistant made by Moonshot AI, a company based in Beijing. It runs on Kimi K3.1 Pro, released in late October 2026 with a 1-million-token context window, and focuses on agent tasks: deep research, slides, documents, spreadsheets, websites and coding through Kimi Code. You can start free at kimi.com. The catch is privacy: the policy linked from kimi.com stores personal information in mainland China.
| Made by | Moonshot AI, Beijing, China |
|---|---|
| Free plan | Yes, on web and mobile. Kimi Code needs a paid plan |
| Models | Kimi K3.1 Pro and K3.1 Pro Highspeed; Kimi K3 in the API |
| Best at | Agent tasks: research reports, slides, documents, spreadsheets, websites, coding |
| Shows sources | Has search and a Deep Research agent; check each report's links before you rely on it |
| Team plans | Kimi Business and Kimi Enterprise, with admin controls |
| Your data | Stored in mainland China under the kimi.com policy; de-identified content may train models unless you ask support to stop |
What Kimi is in late 2026
Kimi started as a chatbot known for reading very long documents. It is now a family of products from Moonshot AI: Kimi on the web at kimi.com and in apps for iOS, Android and HarmonyOS; Kimi Work, a desktop app for Windows and Apple silicon Macs; Kimi Code, a coding agent with a desktop app, a command-line tool and a VS Code extension; a Kimi browser extension; and the Kimi API for developers.
The model behind all of it is Kimi K3.1 Pro, which Moonshot describes as the first model in its K3.1 series and its most capable open-weight model in the 3-trillion-parameter class. It has a 1-million-token context window and arrived in late October 2026 together with K3.1 Pro Highspeed, which Moonshot says runs two to three times faster. Moonshot said the full model weights would be released by 4 November 2026.
On Terminal-Bench 4.0, a test of coding agents working in a terminal, Moonshot reports 65.7% for K3.1 Pro, 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. These are Moonshot's own figures, and several of its other benchmarks are internal, so treat them as a reason to try Kimi rather than as proof. Our explainer on Kimi K3.1 Pro and the K2 cutoff covers the numbers in more detail.
What Kimi is best at
Agent tasks that produce a finished file
Kimi's help centre is organised around agents rather than chat: Agent mode for multi-step tasks, Agent Swarm for running several agents in parallel, Deep Research for long reports, and dedicated agents for slides, documents and spreadsheets. If your goal is a finished deliverable, such as a ten-slide market overview or a spreadsheet model, Kimi is designed for exactly that.
Coding
Kimi Code puts K3.1 Pro in your terminal, editor or desktop, and connectors for GitHub, Cloudflare, Vercel and Supabase let it work with the tools developers already use. Moonshot says K3.1 Pro also builds games and 3D web experiences with engines such as Unity, Unreal Engine and Godot.
Financial research and modelling
Moonshot pitched K3.1 Pro at finance work and reports a score of 64.4 on OfficeQA Pro v2, a test of analysis over real financial documents. It also says the model's claim-level hallucination rate on its hardest internal accuracy test fell from 23.4% for K3 to 8.3%. That is a real improvement on paper, but it is an internal stress test, so keep checking every figure.
3D models and long inputs
3D Make, now live on kimi.com and in the Kimi app, lets you preview and edit 3D models, download them for 3D printing or export them as videos. And the 1-million-token window means a whole codebase or a stack of reports fits in one conversation.
Connected work
Kimi connects to Gmail, Google Calendar, Google Drive, Notion and Obsidian, and supports scheduled tasks, memory and installable skills. Connect only accounts you are comfortable sharing, given the privacy terms below.
Where Kimi falls short
- Your data goes to China. The privacy policy linked from kimi.com says personal information is stored in mainland China, and there is no in-app switch to stop training.
- Self-reported numbers. Most benchmark results come from Moonshot, and several tests are its own. Independent results may differ.
- Plans keep changing. Plan names and allowances have changed during 2026, so check the pricing page before you commit.
- Old models are retiring. Since 31 October 2026, K2 series models are closed to newly registered API users, and Moonshot plans to retire all K2 series models by the end of December 2026.
How to get better results from Kimi
- Start with the right agent. Pick Deep Research, slides, docs or sheets before you type. A request in plain chat gets a plainer result.
- Write a brief, not a question. Give the goal, the audience, the length and the format, plus what a good result looks like.
- Ask for a plan first. For long tasks, ask Kimi to list its steps and wait for your approval, so you can correct the direction before it spends twenty minutes on the wrong thing.
- Demand sources and assumptions. For research and finance, ask for a link for every claim and a list of assumptions in a separate table.
- Use Highspeed for quick turns. Switch to K3.1 Pro Highspeed for short edits and back to K3.1 Pro for hard reasoning.
Privacy settings and where to find them
Read this before you paste anything sensitive. The Kimi privacy policy linked from kimi.com, last updated on 24 August 2026 and effective from 31 August 2026, names Beijing Moonshot Technology Co., Ltd. as the operator and says personal information is stored within mainland China. It says the content you enter and the answers you get may be used to optimise Moonshot's models after encryption and de-identification.
There is no toggle for this. To stop your content being used for training, you contact Kimi customer service through the contact details in section 11 of the privacy policy, and Moonshot says it stops after verifying your identity. Memory is different: the policy says Kimi may remember and summarise details from your chats, and you can switch that off, or view and manage what it saved, on the product's settings page. The terms also say the service is mainly intended for adults, and the privacy policy says users under 18 need a parent's or guardian's consent. For developers, the Kimi API offers Zero Data Retention.
Which plan suits whom
The free plan is enough to test Kimi's agents on real but non-sensitive work. A paid membership makes sense if you lean on the agents every day or want Kimi Code, which is only included in paid plans. Kimi Business and Kimi Enterprise add member management and workspaces, and Enterprise admins must switch K3.1 Pro on for their organisation. Developers pay for the Kimi API separately. Check the official pricing page for current plans and prices.
Pick Kimi if you want an agent that turns a brief into slides, reports or code, and your material is not confidential. Skip it if your company's rules forbid storing data in China. Compare it with others in Kimi vs DeepSeek, our ChatGPT alternatives and the best AI for research, or start the Kimi lessons.
Prompts to copy
Research report with sources
Use Deep Research. Write a 1,000-word briefing on [topic] for [audience]. Use only sources from the last 12 months, give a link and date for every claim, and end with a table of the three things that are still uncertain.
You get: A dated, sourced briefing with open questions flagged
Slides from a brief
Make a 10-slide deck for [audience] on [topic]. Slide 1 states the main recommendation. Each slide has a headline that is a conclusion, no more than 30 words of text and a suggested chart. Before you start, show me the outline and wait for my OK.
You get: An outline to approve, then a finished deck
Spreadsheet model
Build a simple 12-month revenue model in a spreadsheet for [business]. Inputs on one sheet: [price], [customers at start], [monthly growth], [churn]. Outputs on another sheet with a chart. List every assumption in a separate table.
You get: A working model with clear assumptions
Read a long contract
I have uploaded [contract]. List every clause about termination, liability, payment terms and renewal. For each, quote the clause, explain it in one plain sentence and flag anything unusual for a [type of business].
You get: A clause-by-clause summary you can check
Fix a bug in Kimi Code
In this repository, [describe the bug and how to reproduce it]. Find the cause, propose a fix, write a test that fails before the fix and passes after, and explain the change in three sentences. Do not change unrelated files.
You get: A tested fix with a short explanation
Weekly scheduled digest
Every Monday at 8 a.m., search for news from the past week about [industry or company]. Give me five bullet points, each with a link and date, and one sentence on what it means for [my role].
You get: A recurring, sourced news digest