Skip to content
← All guides

How to Use Kimi K3

China just dropped the most powerful open model ever made. It rivals Claude Fable 5 and GPT-5.6, it's #1 in the world for frontend coding, and you can use it free today. Here's the link and exactly what to do with it.

What Kimi K3 actually is

On July 16, a Chinese lab called Moonshot AI released Kimi K3, a 2.8-trillion-parameter model. That makes it the largest open-weight model anyone has ever shipped.

Two things make this a big deal:

1. It's neck-and-neck with the best US models. On independent benchmarks it trades blows with Claude Fable 5 and GPT-5.6, and beats them on a few. It took the #1 spot on the Frontend Code Arena (human-judged coding), beating both. It's #2 in the world on a long-horizon agentic work benchmark, behind only Fable 5.

2. It's open and cheap. The full weights drop July 27, meaning anyone can download and run it. And on the API it's about a third of the price of the US flagships ($3 per million input tokens, $15 output, dropping to $0.30 on cached input).

For most people the takeaway is simpler: there's now a frontier-level model you can use for free, today.

  • 1M
    token context window. Feed it entire codebases or books at once.
  • Vision
    native. It can see images, screenshots, and diagrams.
  • Thinking
    always-on reasoning mode. It thinks before it answers.

The 5 best ways to use it

I pressure-tested it against the stuff I actually do. These are where it shines. Each one comes with a prompt you can copy straight in.

1. Frontend + "vibe coding"

This is its best event. It's ranked #1 in the world for frontend code by human preference, above Claude Fable 5 and GPT-5.6. If you're building a landing page, a tool, or a quick app, this is a genuine reason to open Kimi instead of your usual model.

Copy-paste prompt
Build me a single-page landing site for [your idea]. Modern, clean, mobile-first. One hero with a headline and email signup, three feature cards, and a footer. Give me the full HTML + CSS in one file I can paste and run. No frameworks.

2. Long-horizon agent work (the stuff that takes hours)

K3 is built for tasks that don't finish in one reply. Moonshot's own demo had it autonomously design a working chip over 48 hours straight, reading docs, making decisions, testing, and fixing its own mistakes. You won't do that. But it means it's excellent at multi-step projects that would normally need a dozen back-and-forths.

Copy-paste prompt
I want to [launch a newsletter / build a Notion system / plan a product launch]. Act as an agent: break this into every step, do the parts you can do right now (draft the copy, build the outline, write the templates), and tell me exactly what you need from me to keep going. Don't stop at a plan. Start building.

3. Deep research

It set a state-of-the-art score on BrowseComp, the hardest “go find the answer” benchmark. In one case it reproduced a physics calculation that takes a senior researcher two weeks, in about two hours, cross-checking 20+ papers on its own. For normal research (a competitor, a market, a decision) it's a beast.

Copy-paste prompt
Research [topic]. Find the 5 most credible sources, pull the key facts from each, flag anything the sources disagree on, and give me a one-page brief a smart person could act on. Cite everything.

4. Feed it something huge

The 1-million-token context window is the biggest sleeper feature. Dump in a whole codebase, a full book, a year of meeting notes, or a giant PDF, and ask questions across all of it at once. No chunking, no “sorry, that's too long.”

Copy-paste prompt
Here's [a long document / transcript / codebase]. Summarize it, pull out the 10 most important points, list every open question or decision left unresolved, and tell me what I should do next.

5. Everyday knowledge work

It ranked 3rd in the world on a benchmark that measures real tasks across 44 jobs, ahead of Claude Opus 4.8. Translation: for the normal work most people do with AI (writing, analyzing, planning, emails, summarizing), it's flat-out one of the best options available, and it's free.

Copy-paste prompt
Here's a rough dump of my thoughts on [thing]. Turn it into [a clean email / a structured plan / a post]. Match a direct, no-fluff tone. Ask me anything you need before you write.

For the developers

Moonshot also ships Kimi Code, an open-source coding agent in your terminal, basically their answer to Claude Code. It defaults to their models but works with others, plugs into VSCode, Cursor, and Zed, and it's free and open-source.

If you live in the terminal, it's worth a look: github.com/MoonshotAI/kimi-cli

You don't have to switch everything

You don't need to move your whole life over to Kimi K3. Keep using what works for you.

But the story here is bigger than one model. An open model, one anybody can download and run for free, just caught the best closed models in the world. That gap used to be six months or more. It's basically gone.

The people who win in this era aren't the ones loyal to one tool. They're the ones who stay curious, try the new thing the week it drops, and keep the best few in their back pocket.

So go try it. It's free. It takes 30 seconds. kimi.com