Kimi AI is a smart assistant from Moonshot AI. Ask it questions, hand it long files, or let it build documents, slides, spreadsheets and websites for you. This guide explains what Kimi can do, what it costs, and how to get started, in plain words.
Kimi AI is a smart assistant you can chat with. You type a question or give it a job, and it writes back in plain language. It can also read your files, search the web and build things for you.
Kimi is made by Moonshot AI, a company in Beijing founded in 2023. The first Kimi chatbot came out in October 2023. Its newest and most capable model is Kimi K3, released in July 2026.
Example: you ask, Kimi reads the file, then gives you a summary, slides and a chart.
Six plain reasons people choose Kimi, and a few things worth knowing before you rely on it.
You can chat with Kimi K2.6 on the web or your phone at no cost. No credit card is needed.
See the free plan →Kimi K3 can take in about one million tokens at once. That is enough for a whole codebase or a few long reports together.
PDF summaries →In agent mode Kimi plans the steps and carries them out. Agent Swarm splits a big job between many helpers working at once.
How agent mode works →Documents, slides, spreadsheets, websites, research and code all live in the same app, so you do not need a separate tool for each.
All features →On the API, Kimi K3 is $3 in and $15 out per million tokens. The list price for GPT-5.6 Sol is $5 and $30, and for Claude Fable 5 it is $10 and $50.
API pricing →Moonshot publishes the model weights. You can run Kimi on your own machines, or use it through Amazon Bedrock and GitHub Copilot.
Kimi K3 guide →Want a fuller verdict? Read our Kimi AI review or the comparison with other assistants.
New to Kimi? Follow these five steps. It takes about two minutes and costs nothing to try.
Go to kimi.com in your browser, or install the Kimi app on your Android phone or iPhone. Always check the address says kimi.com.
How to check you are on the real site →Choose a sign-up option on the page and follow the prompts. You do not need a credit card for the free plan.
Login guide →Click the message box, write your question or task in everyday words, and press Enter. Kimi answers in a few seconds.
Attach a PDF, document or image if you want Kimi to read it. Use a quick mode for simple questions, a thinking mode for hard ones, and Agent mode when you want Kimi to carry out a whole task.
What agent mode does →Read the reply and ask a follow-up, such as "make it shorter" or "explain step 2 again". Kimi can make mistakes, so check important facts yourself.
Kimi works in any web browser with nothing to install. If you prefer an app, there is one for Windows and Mac computers and one for Android and iPhone. You sign in with the same Kimi account on each.
One Kimi account works on all four kinds of device.
On a computer, that is kimi.com. On a phone, it is your phone's own app store.
Pick the version for your device and follow the prompts. It takes a minute or two.
Use your Kimi account, or create one for free. The same account works on every device.
Needs: Windows 10 or later
Needs: macOS 12 or later, on a Mac with an Apple chip (M1 or newer)
Needs: Android phones and tablets
Needs: An iPhone with the Apple App Store
Just open kimi.com in Chrome, Edge, Safari or Firefox and start chatting. This is the quickest way to try Kimi, and it works on Linux and Chromebooks too.
The desktop app includes a Work mode, called Kimi Work. It can help with files on your own computer. It is still in beta and uses membership credits.
Kimi Code is a separate tool you install from the terminal.
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bashMac and Linux, above. Windows PowerShell, below.
irm https://code.kimi.com/kimi-code/install.ps1 | iexYes. Kimi AI has a free plan, and you do not need a credit card.
The free plan is called Adagio. It lets you chat with the Kimi K2.6 model on the web and on your phone at no cost. Kimi's most powerful model and its bigger tools need a paid plan, which starts at $19 a month.
| Feature | Free (Adagio) | Moderato |
|---|---|---|
| Price | $0 | $19 a month |
| Credit card needed | No | Yes |
| Chat with Kimi K2.6 | Yes. Does not use credits | Yes. Does not use credits |
| Kimi K3, the top model | Little or none. It is made for paid plans | Yes |
| Agent tasks each month | About 6 | About 60 |
| Tasks running at once | 1 | 2 |
| Agent Swarm | No | Yes. 25 uses a month |
| Kimi Code | No | Yes |
| Professional data lookups | 200 a month | 2,000 a month |
Figures follow Kimi's help centre, checked on 8 October 2026. "Agent tasks" is Kimi's own rough estimate of how far the monthly credits go.
There are four: $19, $39, $99 and $199 a month. Each one gives more credits and more features. Paying yearly saves about 20%.
The API for developers has no free tier. You top up at least $1 and pay for what you use. Once your top-ups reach $5, Moonshot adds a $5 voucher.
The Kimi K3 model is free to download, but it needs very large and costly computers to run. For most people a plan or the API is far cheaper.
Kimi has two separate price lists. Most people only need the first one. Prices are in US dollars and were checked on 8 October 2026.
For chatting and using Kimi's tools on the web or your phone. You pay a fixed amount each month, or nothing on the free plan.
| Plan | Per month | Per year | Best for |
|---|---|---|---|
| Adagio | $0 | Free | Trying Kimi and everyday chat |
| Moderato | $19 | $180 | Regular use with Kimi K3 and agents |
| Allegretto | $39 | $372 | Power users: Kimi Claw and faster coding |
| Allegro | $99 | $948 | Heavy daily use |
| Vivace | $199 | $1,908 | The highest limits |
Paying yearly saves about 20%. Chatting with Kimi K2.6 is free on every plan. See what each plan includes.
For building Kimi into your own app or tool. There is no monthly fee. You top up a balance (from $1) and pay only for what you use, counted in tokens.
| Model | Input | Output | Best for |
|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | Hardest jobs |
| Kimi K2.7 Code | $0.95 | $4.00 | Coding |
| K2.7 Code HighSpeed | $1.90 | $8.00 | Fast coding |
| Kimi K2.6 | $0.95 | $4.00 | General use |
Prices are per one million tokens. A token is a small piece of a word, about three to four letters of English. Full API price list.
A plan and the API are billed separately. Paying for one does not include the other. Plan or API: how to choose.
Say you send Kimi a long document (100,000 tokens) and get back a detailed answer (20,000 tokens).
Kimi's thinking counts as output, so hard questions cost a little more. API cost guide.
Agent credits are Kimi's approximate task-equivalents of a token-metered pool. ✦ Kimi's help centre describes K3's full 1M-token window as a higher-tier benefit - confirm the tier on kimi.com before buying for context alone. Kimi Code also has its own 5-hour and weekly rate limits. Mainland China plans are priced separately (about ¥49-¥699 a month). Full pricing details →
A "model" is the brain behind Kimi. Moonshot AI has released a new one every few months, and each is better at certain jobs. Here is the short version: how they arrived, and which one to pick today.
One year of Kimi models: from Kimi K2 in July 2025 to Kimi K3 in July 2026.
Older models: Kimi K2.5 and Kimi K2 are no longer offered on Moonshot's API, but their open weights can still be downloaded. Still unsure? See how to choose a Kimi model.
Kimi K3 is Moonshot AI's newest and most capable model. It came out on 16 July 2026, and its weights were opened to everyone on 27 July. Here is what that means in plain words.
The size of its "brain". It is the largest model anyone has released openly.
How much it can read at once: a whole codebase or several long reports.
It understands text, images and video in the same conversation.
API price for input and output. Less than the top US models.
Running K3 yourself needs very large hardware. Moonshot recommends 64 or more accelerators.
Terminal-Bench 2.1 score out of 100. Higher is better.
SWE-Marathon score out of 100. Higher is better.
Scores are from Moonshot's own Kimi K3 model card, with every model at its highest effort setting. Each model was run with different tools, so treat small gaps as a tie. See the full comparison table.
Three more things Kimi can do that are worth knowing about.
About 180 tokens per second on typical coding inputs and up to 260 tok/s on short contexts - roughly 5-6× standard K2.7 Code. Same model, faster serving. Available on the API and to Kimi Code members on Allegretto and above.
K2.6 introduced swarms of up to 300 sub-agents and 4,000 coordinated steps; K3 Swarm runs in the Kimi apps on membership credits. Agent Swarm (beta) starts at Moderato with 25 uses a month and 2 concurrent subtasks, rising to 240 uses and 8 concurrent subtasks on Vivace. More in the Agent Swarm guide and Agent Mode guide.
Kimi's professional database covers World Bank datasets, market data and research, with 200 calls a month on the free plan up to 24,000 on Vivace. Since September 17, Kimi for financial services adds S&P Global Market Intelligence, Crunchbase, Wind, Tianyancha and FRED sources.
What do people actually use Kimi for? Find the group closest to you, see what it can help with, and copy the example to try it yourself.
"Explain photosynthesis as if I am 12, then give me five quiz questions."
"Write three friendly Instagram captions for a new coffee blend."
"Find out why this function returns null and write a test for it."
"Research the e-bike market in Europe and give me a short brief with sources."
"Turn this report into a ten-slide deck for managers."
"Compare the last four quarters of revenue for these two companies in a table."
"Build a one-page website for my dog-walking business with a contact form."
"Check 50 competitor websites and list each one's prices in a spreadsheet."
The examples are starting points, and results will vary. Kimi can get facts and numbers wrong, so check anything important before you use it. Agent features, Kimi K3 and some data sources need a paid plan. See membership pricing.
A prompt is simply the message you type to Kimi. The clearer the prompt, the better the answer. Here is an easy way to write one, plus ready-made prompts you can copy.
You are a friendly tutor. Explain how rain forms. It is for a 10-year-old. Use 5 short steps.
You do not need all four parts every time. For a quick question, the task alone is fine. Add the others when the first answer misses the mark.
The four parts drop into place, the prompt is sent, and Kimi replies with a clear answer.
Press "Copy prompt", paste it into Kimi, and replace anything in [square brackets] with your own words.
Summarize the attached file in 5 bullet points. Then list any dates, numbers and action items in a table.
Explain [topic] in simple words, as if I am new to it. Give one everyday example and end with 3 quick quiz questions.
Write a polite email to [person] about [subject]. Keep it under 120 words and end with a clear next step.
Improve the text below. Fix the grammar, make it clearer and keep my tone. Then list the changes you made.
[paste your text]
Here is my code and the error message. Explain what is wrong in plain words, show the fixed code, and tell me how to test it.
[paste code and error]
Research [topic] using recent sources. Give me a one-page brief with the key facts, two opposing views and links to the sources.
Turn the attached document into a 10-slide presentation for [audience]. Give each slide a title, 3 bullet points and a speaker note.
My spreadsheet has these columns: [list columns]. Write a formula that [what you need] and explain how it works step by step.
Plan and carry out this task step by step: [task]. Show me your plan first and wait for my OK before you start.
Deep Research is Kimi's research helper. You ask one question. Kimi then searches the web, reads the pages for you, and comes back with a long written report that links to its sources.
Open Deep Research on kimi.com, or switch to Deep Research mode in the app, and type what you want to find out.
Kimi checks what you mean before it starts. Reply briefly, or choose "Include everything" to skip.
It plans searches, reads pages and checks its findings. This takes a while, so you can close the page. Kimi tells you when it is done.
You get a structured report with clickable sources. You can save it as a PDF or Word file, or open a visual web version to share.
Kimi finds many pages, keeps only the best ones, and turns them into a report with sources.
How many times it stops to reason during one task, on average.
Pages it finds per task, on average. It keeps only the best few.
Sources linked in a typical report, so you can check them.
How long one report usually takes. You can leave and come back.
These are averages published in Kimi's own help centre. Your results will vary with the question.
Deep Research uses your membership credits. One report takes roughly 5 to 10% of a month's credits on the Moderato plan ($19). Credits are returned if a task fails.
Got a long PDF and not much time? Upload it to Kimi and ask for the short version. Kimi reads the whole file, picks out what matters, and can answer your questions about any part of it.
Go to kimi.com or open the Kimi app and start a new chat.
Use the attach button next to the message box and choose the file from your device.
For example: "Summarize this in 5 bullet points." Say how long and how detailed it should be.
Dig deeper with questions such as "What does it say about costs?" or "Explain section 3 again."
Kimi finds the key lines, turns them into a short summary, and answers follow-up questions about the file.
Get the main findings of an annual report or white paper without reading every page.
See what a study asked, how it was done and what it found, in plain words.
Spot the key dates, duties and fees. Use it to prepare, not as legal advice.
Pull figures, dates and names out of the text and into a neat table.
Upload two versions or two offers and ask what is different.
Turn a textbook chapter into notes, flashcards or practice questions.
Attach your PDF first, then paste one of these. Replace anything in [square brackets] with your own words.
Summarize the attached PDF in 5 bullet points. Use simple words and keep each point to one sentence.
Read the attached PDF and list every important date, number and name in a table. Add the page number where you found each one.
Explain the attached paper to someone who is new to the topic. Say what question it asks, how it was studied and what it found.
Give me a short summary of each section of the attached PDF, in order, with the section title above each summary.
In the attached PDF, find everything it says about [topic]. Quote the key sentences and give the page numbers.
Compare the two attached PDFs. List what is the same, what is different, and anything that appears in only one of them.
Kimi Sheets is a spreadsheet helper. You describe what you need in everyday words, and Kimi builds a real Excel file with working formulas, pivot tables and charts. You do not need to know any formulas yourself.
One sentence in, a finished spreadsheet out: data, growth figures, a total, a chart and a file to download.
Type your goal, for example "make a monthly sales report". You can also upload a file you already have.
It organizes the data and adds the formulas, tables and charts for you.
Preview the result, ask for changes, then download it. You can keep editing the file in Excel afterwards.
Describe the result you want. Kimi writes real Excel formulas such as VLOOKUP and XLOOKUP, so you do not have to remember them.
It groups and totals large tables, and can change the rows, columns and filters when you ask.
Bar, line, scatter and pie charts are made straight from your data.
Ask in plain words to remove duplicates, fix formats and tidy up columns.
It can build an interactive dashboard with filters, so you can explore the numbers.
Move between Excel and PDF, Word, PowerPoint, CSV or JSON while keeping the content and layout.
Make a monthly household budget spreadsheet with income, fixed costs and spending categories. Add totals and a pie chart of where the money goes.
Clean the attached spreadsheet. Remove duplicate rows, fix the date format and make every name start with a capital letter. Tell me what you changed.
Use the attached sales data to build a pivot table of sales by region and month. Add a bar chart and list the three best-selling products.
Kimi can build a working website from a description. You say what you want, Kimi designs it and writes the code, and you publish it with one click. Moonshot calls this feature Kimi Build. No coding is needed.
Say who it is for, which pages you need and the style you like. You can also upload a screenshot or design to copy, a screen recording, or start from a template.
Keep chatting to adjust it. Or click any part of the preview, write a note such as "make this button bigger", and send it to Kimi.
Check how it looks on a computer and a phone, then press Publish. You get a web link you can share with anyone.
Kimi builds the page block by block, shows how it looks on a phone, then publishes it with a link.
This list follows Kimi's own help centre.
Published sites get a link that ends in .ok.kimi.link. You can choose the first part, using 3 to 30 lowercase letters, numbers or hyphens.
The link stays online until you unpublish it.
Build a website for my coffee shop. Add a home page with a welcome message, a menu page with prices, opening hours, a map and a contact form. Use warm colours.
Create a portfolio website for a photographer. Show the photos in a grid, add an About page and a contact page, and keep the design clean and minimal.
Build a simple BMI calculator web page. Let people switch between metric and imperial units, and show a short health tip with the result.
These are Kimi's two tools for building software. Kimi Code is a coding helper that works inside your own project. Visual Coding means Kimi can look at a picture of a design and write the code to match it.
A coding assistant you run in the terminal or in VS Code. You tell it what to do in plain words and it does the work in your files.
Kimi can see as well as read. Show it what you want and it writes code that looks the same.
Both tools together: Kimi scans a design, writes the code, runs the tests and gives you a preview.
On Mac or Linux, paste this line into your terminal. Windows and Homebrew options are in the guide.
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bashOpen your project folder, start Kimi Code and type /login. It connects to your Kimi account in a few seconds.
Type what you want, such as "fix the failing test". Use /model to switch between Kimi K2.7 Code and Kimi K3.
Look at the failing test in this project. Find the cause, fix it, run the tests again and tell me what you changed.
Here is a screenshot of a page design. Build it as a responsive web page with HTML and CSS, and match the colours, spacing and fonts as closely as you can.
Read this codebase and explain how it is organized. List the main folders, what each one does, and where I should start if I want to add a new feature.
Kimi has a few more tools besides the ones covered earlier on this page.
Write, convert, review, and translate documents. Upload PDFs, Word files, slides, or spreadsheets - get clean structured outputs with summaries and professional formatting. LaTeX PDF support. See also PDF summaries with Kimi.
Kimi Docs guide →Two creation modes: Adaptive (research-first) and Visual (image-led). Chart parsing converts static chart images to native editable PPTX objects. Agentic Slides converts any document to a deck. Draws on membership credits.
Kimi Slides guide →Always-on cloud agent - no server setup needed. Persistent long-term memory, scheduled automation and an Android companion. Available on Allegretto and above.
Kimi Claw guide →Moonshot's desktop agent for knowledge workers, launched in June 2026 - research and productivity tasks run as agent workflows rather than chat. Shares the membership credit pool; using K2.6 as an agent inside Kimi Work consumes credits even though K2.6 chat is free.
Kimi Work at kimi.com →New since September 17, 2026. Kimi can pull directly from S&P Global Market Intelligence, Crunchbase, Wind, Tianyancha, local financial news providers and FRED for analysis and reports. Named users include CICC and HongShan. No separate interface - it runs inside the regular Kimi apps, with data access varying by plan.
Open kimi.com →The Kimi API lets developers put Kimi inside their own apps, websites and tools. Instead of a person typing in the Kimi app, your program sends a message and gets the answer back.
Your app sends a request, a Kimi model works on it, and the answer comes straight back.
Sign up at platform.kimi.ai, add at least $1, and create an API key. The key is like a password for your program, so keep it private.
If your code already talks to OpenAI or Anthropic, you only change two things: the web address and the model name.
Kimi sends back the answer. You pay only for the text going in and coming out, counted in tokens.
K3, K2.7 Code, K2.7 Code HighSpeed and K2.6.
Top up from one dollar. No monthly fee.
K3 costs more: $3 in and $15 out.
Works with tools built for OpenAI or Anthropic.
kimi-k3kimi-k2.7-codekimi-k2.7-code-highspeedkimi-k2.6OpenAI and Anthropic SDK compatible. Token-based billing, separate from app membership. Prices below are from Moonshot's platform docs as of 8 October 2026 and exclude tax.
kimi-k3 - ✦ Flagship · 1M ctx · reasoning_effort low/high/max
kimi-k2.7-code - Coding specialist · 262K ctx
kimi-k2.7-code-highspeed - ~180-260 tok/s ⚡
kimi-k2.6 - General purpose · thinking on/off
kimi-k2.5 / moonshot-v1-* - retired Aug 31, 2026 (404)
kimi-k2-* (incl. thinking, turbo) - retired May 25, 2026
kimi-latest - Jan 28, 2026 · kimi-thinking-preview - Nov 11, 2025
Cache-hit input: $0.30 (K3) · $0.19 (K2.7 Code) · $0.38 (HighSpeed) · $0.16 (K2.6). K3 also bills cache writes: $3.00 at the default 5-minute TTL, $6.00 at the 1-hour TTL. No long-context surcharge on K3. Batch API jobs cost 60% of the standard price and cover K2.7 Code and K2.6 only - K3 is not supported. Full pricing guide →
POST /v1/tools/search - $0.002 per call
POST /v1/tools/search_pro - $0.003 per call · site + time filters
POST /v1/tools/fetch - $0.002 per call · URL to Markdown
Charged only on successful, non-empty responses. The legacy $web_search built-in tool ($0.005 per call) is expected to be deprecated on October 20, 2026. Also new: usage is deducted 50% from cash and 50% from vouchers until one runs out (contract customers are unaffected).
K3: thinking always on · top-level reasoning_effort = low | high | max (default) · temperature 1.0, top_p 0.95, n and penalties fixed - omit them · images as base64 or ms:// only · video via files API
K2.7 Code: thinking always on
K2.6: thinking switchable (on/off)
K3's default max_completion_tokens is 131,072 (max 1,048,576) - set an explicit lower limit. Rate limits follow cumulative top-up: Tier 0 ($1) allows 1 concurrent request and 3 RPM; Tier 1 ($10) 15 concurrent and 100 RPM; Tier 5 ($3,000) 100 concurrent, 300 RPM and 5M TPM.
Weights: Hugging Face → moonshotai/Kimi-K3. Native MXFP4 weights / MXFP8 activations, ~1.56 TB; Moonshot recommends 64 or more accelerators. Official recipes for vLLM, SGLang and TokenSpeed. Run the Kimi Vendor Verifier before production traffic.
Hosted alternatives: Amazon Bedrock (GA Sep 18 - us. and global. profiles, no video), GitHub Copilot (Fireworks-hosted), and gateways such as OpenRouter and Vercel AI Gateway. Some third parties also sell a faster "K3 Fast" serving tier at a premium (about $4.50 / $22.50 per 1M) - that is a provider offering, not a Moonshot model ID.
How does Kimi stack up against the three assistants people compare it with most? The table gives the facts side by side. The cards under it help you choose.
| Feature | Kimi | Perplexity | DeepSeek | ChatGPT |
|---|---|---|---|---|
| The basics | ||||
| Made by | Moonshot AI, China | Perplexity, USA | DeepSeek, China | OpenAI, USA |
| Best known for | Long tasks, agents and open models | Search answers with sources | Very low prices | The all-round assistant |
| Top model | Kimi K3 | Its own Sonar, plus OpenAI, Claude and Gemini models on paid plans | DeepSeek V4 Pro | GPT-6 Astra on higher plans. GPT-5.6 Sol on Plus |
| Can you download the model? | Yes. Open weights | No | Yes. Open weights | No |
| Price | ||||
| Free plan | Yes. Kimi K2.6 chat is free | Yes. About 3 Pro searches a day | Yes. The chat app is free | Yes. Uses GPT-5.6 Luna |
| Paid plans, per month | $19, $39, $99 or $199 | $20 Pro or $200 Max | No paid plan for the chat app | $8 Go, $20 Plus, $100 or $200 Pro |
| API price, top model per million tokens, in / out | $3 / $15 | Separate Sonar API | $1.74 / $3.48 | $5 / $30 for GPT-5.6 Sol $4 / $20 on offer until 21 Nov 2026 |
| Research | ||||
| Deep research reports | Yes. Takes 10 to 25 minutes and links about 26 sources | Yes. This is its main strength | Not confirmed in the sources we checked | Yes. Plus includes 10 full runs a month |
| Shows its sources | In Deep Research reports | Yes. On every answer, even free | Not confirmed | In Deep Research reports |
| Free research allowance | Not stated by Kimi | About 1 Deep Research run a month | Not confirmed | 5 lightweight runs a month |
| Tools and apps | ||||
| Reads very long files | About 1 million tokens with Kimi K3 | Depends on the model you pick | About 1 million tokens | About 1 million tokens with GPT-5.6 Sol |
| Built-in tools | Docs, Slides, Sheets, Websites, Kimi Code, agents | Search, Labs, Computer agent | Chat and file upload | Codex, Agent Mode, ChatGPT Work |
| Where it runs | Web, Windows, Mac, Android, iPhone | Web, Android, iPhone | Web, Android, iPhone | Web, desktop, Android, iPhone |
On a small screen, swipe the table sideways to see every column.
Checked on 8 October 2026. Kimi details come from Moonshot's own pages. Perplexity, DeepSeek and ChatGPT details come from independent pricing trackers, and those companies change plans often, so check their sites before you pay. "Not confirmed" means our sources did not say either way.
Pick Kimi if
Pick Perplexity if
Pick Kimi if
Pick DeepSeek if
Pick Kimi if
Pick ChatGPT if
Rebuilt for October 2026 from the Kimi-K3 model card (all models at maximum reasoning effort) and current official price lists. Newer rivals have shipped since K3's card was published - see the note under the table.
Kimi K3 ✦ Moonshot AI · Jul 2026 |
Kimi K2.7 Code Moonshot AI · Jun 2026 |
Kimi K2.6 Moonshot AI · Apr 2026 |
GPT-5.6 Sol OpenAI |
Claude Opus 4.8 Anthropic |
Claude Fable 5 Anthropic |
GLM-5.2 Z.AI |
|
|---|---|---|---|---|---|---|---|
| Architecture & Access | |||||||
| Open weights | ✓ Kimi K3 License | ✓ Modified MIT | ✓ Modified MIT | ✗ | ✗ | ✗ | ✓ |
| Parameters (total / active) | 2.8T / 104B | 1T / 32B | 1T / 32B | Undisclosed | Undisclosed | Undisclosed | 753B |
| Context window | 1M (1,048,576) | 262K | 262K | ~1.05M | - | 1M | - |
| Multimodal input | ✓ image + video | ✓ MoonViT | ✓ | ✓ | ✓ | ✓ | - |
| Benchmarks (Kimi-K3 model card · max effort) | |||||||
| Terminal-Bench 2.1 | 88.3 | - | - | 88.8 | 84.6 | 88.0 | 82.7 |
| DeepSWE | 67.5 | - | - | 73.0 | 59.0 | 70.0 | 46.2 |
| FrontierSWE | 81.2 | - | - | 71.3 | 66.7 | 86.6 | 67.3 |
| SWE-Marathon | 42.0 ★ | - | - | 39.0 | 40.0 | 35.0 | 13.0 |
| Toolathlon-Verified | 76.5 | - | - | 74.9 | 76.2 | 77.9 | 59.9 |
| MCPMark-Verified | 94.5 ★ | 81.1 (at launch) | ~73 | 92.9 | 76.4 | 87.4 | - |
| BrowseComp | 91.2 ★ | - | - | 90.4 | 84.3 | 88.0 | - |
| GPQA Diamond | 93.5 | - | - | 94.1 | 91.0 | 92.6 | 91.2 |
| GDPval-AA v2 (Elo) | 1,686 | - | - | 1,736 | 1,593 | 1,747 | 1,510 |
| Video-MME (w. subtitles) | 90.0 ★ | - | - | 89.5 | 86.0 | - | - |
| LMArena Frontend Code Arena | #1 at launch · 1,679 | - | - | - | - | - | - |
| Artificial Analysis Index v4.3.2 | 44 | - | - | - | - | - | 45 (GLM-5.3) |
| Speed & Cost (official list prices · Oct 2026) | |||||||
| API input $/1M | $3.00 | $0.95 | $0.95 | $5.00 (promo $4.00) | $5.00 | $10.00 | $1.40 |
| API output $/1M | $15.00 | $4.00 | $4.00 | $30.00 (promo $20.00) | $25.00 | $50.00 | $4.40 |
| Cached input $/1M | $0.30 | $0.19 | $0.16 | $0.50 | - | $1.00 | - |
| Output speed | ~40 tok/s (first-party) | 260 tok/s ⚡ (HS) | - | - | - | ~60-70 tok/s | - |
| Platform & Ecosystem | |||||||
| Agent Swarm (parallel agents) | K3 Swarm (app) | via K2.6 | 300 agents | Limited | Limited | Limited | Limited |
| In GitHub Copilot picker | ✓ Since Aug 6 (Fireworks) | ✓ Since Jul 1 (Azure) | - | ✓ | ✓ | ✓ | - |
| On Amazon Bedrock | ✓ Since Sep 18 | - | - | ✓ | ✓ | ✓ | - |
| Integrated office tools | Full suite | Docs/Slides/Sheets/Code | Full suite | Canvas | Artifacts | Artifacts | - |
| Free tier | Small credit allowance | Via Kimi Code plans | ✓ Free in chat | Limited | Limited | Limited | Limited |
★ = best in this table. Benchmark rows come from Moonshot's Kimi-K3 model card, which mixes its own runs with third-party leaderboards and uses different agent harnesses per model - read its footnotes before relying on a single row. K2.7 Code and K2.6 figures are Moonshot-reported at their own launches. GPT-5.6 Sol promotional pricing runs at least through November 21, 2026. Not shown: models released after K3's card, including Claude Opus 5 (July 24), GLM-5.3 (August), GPT-6 Astra (September) and Reflection Beam (October 5). Always verify current benchmarks and pricing on official pages.
What has changed in the four weeks since our last update - new places to run K3, new products, the IPO track, and a growing list of allegations and investigations. Checked against Moonshot's platform docs, the Kimi-K3 model card, AWS and major press as of 8 October 2026.
Distribution. On September 18 AWS made Kimi K3 generally available on Amazon Bedrock (moonshotai.kimi-k3, with US and Global cross-Region inference profiles). AWS says prompts stay inside its boundary and are not shared with Moonshot, and that K3 is the first open-weight model on Bedrock with explicit prompt caching. Video input is not supported there. A day earlier Moonshot launched Kimi for financial services, wiring Kimi to S&P Global Market Intelligence, Crunchbase, Wind, Tianyancha and FRED data.
Money. Bloomberg reported on October 6 that Moonshot closed its final private round at about $50 billion and is heading for a Hong Kong IPO in Q1 2027. Reuters had reported a confidential filing on September 3 (about $3B raise; Goldman Sachs, CICC and Deutsche Bank). Moonshot is targeting $2B annualized revenue by year-end, double its reported August run rate.
Scrutiny. Anthropic (Sep 10) and OpenAI (Sep 30) both published allegations tying Moonshot-linked accounts to unauthorized distillation; The Information reported on Sep 22 that China's cyberspace regulator is examining Moonshot and DeepSeek over user data reaching Claude. These are allegations and reports - no sanctions, Entity List action or regulatory findings have been announced. Details in the timeline.
Bloomberg, citing people familiar: the round is closed at roughly $50B (up from $35B in July) and early-look investor meetings could start this month. SCMP has also reported Moonshot is exploring a parallel Shanghai STAR Market listing. Figures are unconfirmed by the company.
The Nvidia-backed US lab's first open-weight model pitches efficiency. Its own table gives Beam 80.1 on Terminal-Bench 2.1 versus 88.3 for Kimi K3.
Fox News reported that researcher Peter Garrigan was able to manipulate a Kimi model into producing dangerous instructions. Moonshot is investigating and in contact with him. Garrigan said US models share the same underlying weakness.
OpenAI said it disrupted a July campaign to extract its models' hidden reasoning - 16,000 requests from 4,000+ users on July 24-25 - and attributed "a core cluster" to individuals associated with Moonshot. It said it could not tie every operator to one actor. Moonshot had not responded publicly.
The Information reported the Cyberspace Administration of China is looking at whether Chinese users' data reached Anthropic's Claude. Regulators have not confirmed a probe or announced any finding.
Serverless, via us.moonshotai.kimi-k3 and global.moonshotai.kimi-k3 profiles; Standard, Priority and Flex tiers; explicit prompt caching. No video input on Bedrock.
Built into the regular Kimi apps rather than a separate interface. Data partners include S&P Global Market Intelligence, Crunchbase, Wind and Tianyancha, plus direct FRED access. Moonshot named CICC and HongShan among users. Data access varies by subscription tier (CNBC).
New REST endpoints /v1/tools/search, /v1/tools/search_pro and /v1/tools/fetch ($0.002-$0.003 per successful call). API usage is now deducted 50% from cash and 50% from vouchers. The legacy $web_search built-in tool is expected to be deprecated on October 20, 2026.
Bloomberg: that would double the reported August run rate of over $1B. TechCrunch noted K3 was generating as many as 300B tokens a day on OpenRouter.
Anthropic's threat-intelligence report alleges Moonshot forwarded nearly 300,000 Kimi customer requests to Claude over about ten days via 5,380 fraudulent accounts, and attributes 23M+ exchanges to Moonshot between May and July. Seven Chinese companies were named. Moonshot declined to comment to Bloomberg; Chinese officials rejected the claims.
NSA, FBI and CISA joint advisory AA26-251A named six China-based labs including Moonshot. Sanctions and Entity List action were floated in July; none has been imposed.
K3 launched Jul 16; open weights Jul 27 under the Kimi K3 License; $35B round Jul 30; K3 in GitHub Copilot Aug 6; Reuters reported revenue-share talks with Microsoft, Amazon and Google Aug 26; kimi-k2.5 and moonshot-v1 retired Aug 31. Hugging Face →
Most problems with Kimi are quick to fix. Start with the six steps below. If that does not work, find your problem in the list and follow the tips.
Reload the tab, or close the app fully and open it again.
Open another website. If that fails too, the problem is your connection.
Use Chrome, Edge or Safari. Turn off ad blockers and privacy add-ons for kimi.com.
This clears a stuck session. Make sure you are on kimi.com.
Very long chats can slow down or act oddly. A fresh one often fixes it.
At busy times Kimi can be slow or queue tasks. Try again in a few minutes.
/login again. Sign-in on a device expires if it is not used for 30 days.| Code | What it means | What to do |
|---|---|---|
401 | Your key is wrong or for the wrong region | Check the key. Global accounts use api.moonshot.ai. Mainland China accounts use api.moonshot.cn. |
404 | The model name does not exist | Older names such as kimi-k2.5 and moonshot-v1 were retired on 31 August 2026. Use kimi-k3, kimi-k2.7-code or kimi-k2.6. |
429 | You are sending requests too fast | New accounts allow only 3 requests a minute. Limits rise as you top up. Slow down and retry. |
400 | Something in the request is not allowed | With Kimi K3, leave out temperature and top_p. Send images as files, not web links. |
Kimi K3 also needs at least a $1 top-up before it will answer. API details.
Use the User Feedback option inside Kimi. Say what you were doing, add a screenshot, and include any version number shown.
We are an independent guide and cannot reach your account. Only Moonshot's own support can.