An independent guide to Kimi
Kimi AI:
what it is and where to start.
Kimi is an AI assistant developed by Moonshot AI. It combines chat, web research and file handling with tools for creating images, websites, presentations, documents and spreadsheets. You can use the assistant directly or access supported models through a separate developer API.
ChinaAI is an independent publication, not Moonshot AI. These buttons open Kimi’s own services.
By China AI Chat Team · Evidence guide updated September 25, 2026 · Tests: September 18–25, 2026 · Our research method
What can Kimi AI do?
Kimi can answer a question in chat or use tools to produce a file, image or website. Start with the result you need. The official overview and product directory describe its tools; availability and allowances depend on your service and account.
Use this table to choose a workflow, then check the findings below before you rely on the output.
| Your task | Start with | What to ask for |
|---|---|---|
| Read or summarize a PDF | Kimi PDF guide | Answers tied to the supplied pages, with missing information flagged. |
| Research a question | Deep Research | A defined question, date range and sources you can open and verify. |
| Prepare office files | Slides, Docs or Sheets | The intended audience, supplied facts or data, and the file format you need. |
| Create or edit an image | Kimi Design | The subject, dimensions, exact text and details that must survive an edit. |
| Build a website prototype | Export code or publish a website | The pages and working controls; specify whether you need files, a public link or both. |
| Work with local files | Kimi Work | A small folder, the files it may read or change, and a named output. |
| Change an existing codebase | Kimi Code | A defined change and acceptance checks you can run independently. |
| Integrate a model into an app | Kimi API guide | The provider and model, usage limits and server-side credential storage. |
For a simple explanation or rewrite, start in the official chat with your source text, intended reader and requested length. A small task whose result you can verify is a useful first trial.
How to use Kimi for your first task
- Open the official assistant. Use kimi.ai and follow the account steps shown there. The official download page directs you to the available desktop and mobile products.
- Choose the result you need. Start in chat for a question. For a website, slide deck or spreadsheet, use the relevant tool available in your account.
- Provide context and constraints. Include the source material, intended audience, desired output and anything that must remain unchanged. Ask Kimi to flag missing information.
- Review the result before using it. Check the source of a factual claim, inspect a generated file, or try the controls in a website. Ask for a targeted revision when a requirement is missed.
Here is an English prompt template for a small document task. It is a suggested starting point, not an additional test result:
Turn the notes below into a one-page project brief in English. Use these headings: objective, audience, deliverables, deadlines and open questions. Preserve the supplied dates and numbers. Do not invent missing details; list them under open questions. Notes: [paste non-sensitive notes here]
For current information, ask for dated sources and check them yourself. Kimi’s current help says it decides whether to search online based on the question. Do not assume every answer used a fresh web search.
Is Kimi AI free?
Kimi offers free access with usage limits, alongside paid memberships. Its credit guide describes free-tier users and explains that multiple features draw from a shared credit pool. A short chat, a research task and a generated website need not consume the same amount.
Check the allowance, supported features, currency and billing period shown in your account before paying. Regional offers differ, and our test account’s Moderato label does not establish what a new account receives today.
The Open Platform API is billed separately. Kimi’s billing and authentication guidance says consumer membership benefits do not become Open Platform balance. Kimi Code credentials and Open Platform API keys are also not interchangeable.
Our Kimi pricing guide explains membership versus API costs, billing periods and token-based examples. Use it to identify which product you are budgeting for.
What our Kimi tests found
The guides show specific results and mistakes, with prompts, inputs, account settings and output evidence. They cover different workflows, so their results are not an overall accuracy score.
The following file, research, image, Work and Code examples were tested on September 23, 2026; the timer website was tested September 18–19. They used a Moderato account. Modes differed, including Instant High for PDF reading and K2.8 Preview for Work and Code; do not assume every result came from the same model or is available on a free account.
PDFs and research: check completeness as well as citations
In our three-file PDF trial, the text PDF answer missed important details; the clean image-only brief and order table passed their checks. Check each answer against its source: this result does not mean scanned PDFs generally work better.
Our Webb chronology report included useful NASA and ESA sources and distinguished related events, but also added a false verification note. Open the cited source and check that it supports the particular sentence, not merely the same topic.
Slides, spreadsheets and documents: inspect the file you will use
- Slides: the five-slide briefing retained the specified facts in the editor. A 25 September follow-up downloaded the same deck as a PPTX with 50 native text objects and checked all five online slides and local rendered previews. Manual editing in desktop PowerPoint remains untested; the guide records the failed attempts and successful export settings.
- Sheets: the original eight-order workbook contained correct formulas and stored totals. We recalculated copies with a separate tool: quantity and cancellation changes updated correctly, but an order added below the unchanged range was left out of the summary. See the exact changes and results. Desktop Excel/LibreOffice behavior remains untested.
- Docs: the downloaded Word brief retained the supplied facts and native tables. The first file split a section across a page break. A 25 September follow-up kept the section together in the online preview while preserving the text and numbers; desktop Word rendering remains untested.
A good-looking editor preview, a correct calculation and a usable exported file are separate checks. The later slide comparison below used a different brief and found errors in the first deck.
Images: an edit can preserve the layout and reduce readability
The first poster in our Kimi Design test kept the requested text and square format. A background edit broadly preserved its layout but made pale lettering harder to read. Inspect the downloaded image after each change, especially text contrast and details you wanted to keep.
Work and Code: review actions and verify the result yourself
In our Kimi Work folder task, the agent preserved all three input files and flagged conflicting times. The action log also showed an unrequested read-only shell command. The prompt did not explicitly prohibit shell commands, so this observation is not proof of a permission bypass; it shows why file instructions and action review matter.
In the separate Kimi Code trial, one approved edit moved a small project from five to seven passing checks. Kimi could not run its own test command in the restricted environment; we verified the result independently. A correct edit and successful end-to-end automation are different findings.
Websites: exporting, saving a draft and publishing are separate
We exported a four-file Focus Break Timer and checked its controls locally. In a separate publishing trial, the public site changed only after Update; saving the draft alone did not change it. Unpublish removed public access, and the demo remains unpublished.
This small static project did not test accounts, databases, payments or production security. Its verified export is also a different result from the code-viewer recovery documented in our later Kimi–ChatGPT comparison.
Compare Kimi with ChatGPT, Qwen and DeepSeek
Choose a study whose task resembles your own. These comparisons used different prompts, settings and plans; combining them into an overall model ranking would be misleading.
Kimi vs ChatGPT: sources, data and a revised web tool
Both returned correct factual judgments and CSV results, while each revised web tool had a keyboard-focus defect. The evidence distinguishes verified file downloads from Kimi HTML recovered through its code viewer.
September 24–25, 2026 · Kimi Moderato/K3 High versus ChatGPT Pro · Different paid-plan conditions.
Kimi vs Qwen: a report, revised quote and slide drafts
Kimi followed the source report more closely, both reconciled the quote and both first decks needed corrections. The quote used page-labelled text, not a PDF upload; an automatically activated Kimi account skill also limits base-model conclusions.
September 24, 2026 · International consumer apps · First outputs retained.
DeepSeek vs Kimi: app workflows and separate API tests
Three official-app tasks checked a changed NASA mission, a revised Word handout and a chart. Kimi produced a Word file; the observed DeepSeek chat route provided copy-ready text. The article labels its separate API investigation so those results are not confused with the app tests.
September 24, 2026 · Official-app observations and separately reported API records.
Still choosing which pair to investigate? Our Chinese AI comparison guide organizes the available studies by task and the evidence you can inspect.
Kimi, Kimi Code and Kimi models: what is the difference?
- Kimi assistant: the consumer service where you chat, upload material and use the creation and research tools available to your account.
- Kimi Work: a desktop agent for working with local files and web-based tasks. Review the access it requests before connecting your work.
- Kimi Code: a coding assistant for a terminal or supported editor, designed to work with a codebase. It is a different workflow from generating a website in the browser.
- Kimi models: named releases such as Kimi K3 and Kimi K2.6. A model is the underlying system; the surrounding application supplies tools, file handling and other product behavior.
- Kimi Open Platform: the developer service for calling supported models from your own software. An API call does not automatically recreate the assistant’s complete website-building or presentation workflow.
These distinctions follow Kimi’s product descriptions and official API model list. Check the exact service and version when comparing capabilities; the model label alone does not tell you which tools or limits an account includes.
Use our Kimi model guide for versions and supported inputs, or the Kimi API guide to choose the correct developer platform and set up an integration.
Where you still need to check the result
- Research and documents: verify citations, dates, calculations and whether the output actually uses the supplied material. We have not measured Kimi’s accuracy across long documents or research topics.
- Generated applications: test behavior beyond the preview. Kimi’s Websites FAQ explains that downloading code does not automatically transfer Kimi-managed login or database data into a standalone deployment.
- Private material: review the privacy policy and account controls in the service you use before uploading confidential files. Our website exercise was not a security or privacy audit.
- Public sharing: treat publishing as a separate decision. In our test, the published site was accessible without signing in. Removing the public page later does not prove that earlier copies were erased.
For evaluating Kimi on your own work, define success before you prompt it: required facts, file format, working controls or tests that must pass. That gives you a concrete basis for deciding whether the output is useful.
Common questions about Kimi AI
Who made Kimi?
Kimi is developed by Moonshot AI, a Chinese AI company. Its official company page provides company information and links to Kimi. ChinaAI is independent of both.
Can I use Kimi in English?
Yes. Our document, research, website and comparison tasks produced English answers to English instructions. Check accuracy for your own subject.
Is kimi.ai or kimi.com the official website?
Both are official destinations: Moonshot’s international company site links to kimi.ai, and Kimi’s help also links to kimi.com. Check the service, region and account terms before assuming their plans or balances are the same.
Is Kimi better than ChatGPT?
Our Kimi vs ChatGPT comparison tested source checking, CSV analysis and a revised web tool on Kimi Moderato and ChatGPT Pro. Both returned the correct factual judgments and CSV results; file delivery and keyboard behavior differed. These limited app tests do not establish an overall model winner.
Is Kimi open source?
Moonshot publishes downloadable model releases through its official model organization. Check the license and computing requirements of the exact release. Downloadable model weights are separate from access to the hosted Kimi assistant and its tools.
Can I chat with Kimi on ChinaAI?
Live visitor chat is not active here. Our chat directory provides official links, a question draft you can copy and clearly labelled recorded examples. A draft is not sent to a model; open the official Kimi service for a new live answer.
Models, API access and pricing
For a model release, use the Kimi model guide. For integration steps, use the Kimi API guide. To separate membership allowances from developer usage costs, read Kimi pricing.
This overview combines provider documentation with separately dated hands-on observations. Core product documentation was reviewed September 21–23; product routes and the evidence summaries were reviewed September 25, 2026. The linked guides preserve their exact test dates, modes, results and limits. Task suggestions are editorial guidance; provider descriptions are not measured performance findings. How we research · Report a correction.