Kimi K3 is Moonshot AI's most capable model, and the largest AI model anyone has released openly. It can read about a million tokens at once, understands images and video, and is built to keep working on a big job for hours with little help. This page explains what it can do, what it costs, where it falls short and how to use it.
The four facts that set Kimi K3 apart from earlier Kimi models.
Moonshot's top model: very large, open, and able to read text, images and video.
Close to the best models from the US on many tests. Moonshot says it still trails the top two overall.
A paid Kimi plan, or the API at $3 in and $15 out per million tokens.
A model is the "brain" behind an AI assistant. Kimi K3 is the biggest brain Moonshot has built. It launched on 16 July 2026, and Moonshot published the model itself for anyone to download later that month.
From Moonshot's Kimi K3 model card and launch post. The page estimate uses about 650 tokens a page. New to these words? See how Kimi works.
Kimi K3 is built for long jobs. You hand over a large coding or research task and it keeps going, using tools and checking its own work, with little supervision.
A million tokens is enough for a whole codebase or several long reports at once, so it can work across them without losing track.
Screenshots, charts, scanned pages and video all work as input. It can look at its own results on screen while it builds something.
Kimi K3 reasons before every answer. You cannot turn that off, but developers can set the effort to low, high or max.
Moonshot's launch post describes several long tasks Kimi K3 completed. These are the company's own showcases, chosen to impress. We have not repeated them.
Built a small compiler from scratch, then used it to train a tiny AI model from start to finish.
Completed a small chip design in a single 48-hour run, using open-source design tools.
Read more than 20 papers and wrote over 3,000 lines of code to reproduce a result, in about 2 hours.
Produced an interactive report covering 42 years of one part of the chip industry.
Cut a teaser from 56 source clips, and made an explainer video about its own design.
Used more than 20 helper agents at once to analyse a set of space-science data.
The everyday versions of these: Deep Research · Agent Mode · Agent Swarm · Kimi Code
Moonshot's model card lists more than 40 tests. Here are eight, picked to show where Kimi K3 leads and where it does not. Higher is better, and the best score in each row is marked.
| Test | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| Terminal-Bench 2.1Coding tasks in a terminal | 88.3 | 88.0 | 88.8 | 84.6 |
| SWE-MarathonVery long software projects | 42.0 | 35.0 | 39.0 | 40.0 |
| FrontierSWEHard software engineering | 81.2 | 86.6 | 71.3 | 66.7 |
| DeepSWEFixing bugs in large codebases | 67.5 | 70.0 | 73.0 | 59.0 |
| BrowseCompFinding hard facts online | 91.2 | 88.0 | 90.4 | 84.3 |
| OSWorld 2.0Using a computer like a person | 58.3 | 66.1 | 62.6 | 55.7 |
| HLE-FullExpert questions, no tools | 43.5 | 53.3 | 44.5 | 49.8 |
| OmniDocBenchReading complex documents | 91.1 | 89.8 | 85.8 | 87.9 |
On a small screen, swipe the table sideways to see every column.
For outside views, see our Kimi AI review and the comparison hub.
Choose Kimi K3 at kimi.com or in the mobile apps. Aimed at paid plans.
The desktop app for longer tasks. Moonshot lists Windows and Apple silicon Macs.
Our earlier research found it offered on Amazon Bedrock. Check your provider's current list.
Download the open weights. Only realistic for teams with very large hardware.
API prices from Moonshot's price page, before tax. Plan details follow our pricing pages and should be confirmed on kimi.com. More: Kimi AI pricing · Is Kimi AI free? · API pricing
Kimi K3 is at its best on big jobs with a clear goal and clear limits. Replace the parts in [brackets].
I have attached [number] documents. Read all of them, then tell me where they agree, where they disagree, and what is missing. Quote the source for each point.
Build [what you want]. Before you start, show me your plan. Do not delete files, install new tools or change [thing] without asking me first.
Here is a screenshot of [app or chart]. Explain what it shows in plain words, then list three things that could be improved.
More ideas: how to use Kimi AI.
Moonshot's launch post is unusually open about where Kimi K3 falls short.
| Kimi K3 | Kimi K2.7 Code | Kimi K2.6 | |
|---|---|---|---|
| Made for | The hardest tasks of any kind | Programming | Everyday general work |
| Reads at once | About 1 million tokens | About 262K tokens | About 262K tokens |
| Size | 2.8 trillion parameters | 1 trillion | 1 trillion |
| Thinking | Always on, three effort levels | Always on | On or off |
| API price, in / out | $3.00 / $15.00 | $0.95 / $4.00 | $0.95 / $4.00 |
| Pick it for | Huge codebases, long research, deep reasoning | Daily coding at a low price | Chat, writing and quick tasks |
On a small screen, swipe the table sideways to see every column.
More: Kimi K2.7 Code · Kimi K2.6 · choose a Kimi model
Most top AI models can only be used through their maker's service. Moonshot published Kimi K3 itself, so other companies and researchers can download it, study it and run it on their own machines.
Businesses can keep their data on their own servers, and are not tied to one supplier.
The weights come under the Kimi K3 License. It is Moonshot's own license, so read it before any commercial use.
Our earlier research put the files at about 1.56 TB, with 64 or more accelerators recommended. This is data-centre scale.
The file size and hardware figures were collected earlier for this site and were not on the pages we read today.