How to use Kimi K3 online, step by step

Kimi K3 is Moonshot's 2.8 trillion parameter reasoner with a one million token context window. The weights are public and almost nobody downloads them, because a model that size lives across a rack rather than on a graphics card.

So the useful question is not how to install it. It is where you type into it. This guide covers both answers we support: the browser Studio, and an OpenAI-compatible endpoint your own scripts can call. The background on the model and the plan it sits in is on the Kimi K3 online page.

Step 1: pick your entry point

Two ways in, and they are not exclusive.

A credit pack needs no subscription. The packs are €5 for 165,000 credits, €10 for 350,000 and €25 for 900,000, all three available on your first purchase, and the credits never expire. A pack carries the whole chat catalog, K3 included. That is the cheap way to find out whether the model suits your work.

The Hosted plan is €19 a month with 900,000 credits, on the same catalog. Upgrading buys volume, not access. All plans and packs are here.

Step 2: get an account and credits

Buy a pack from the credits tab, or start on Hosted if you want the monthly budget. Card details go to Stripe and the payment is secured with 3D Secure.

When the payment lands, the Studio opens on your account. Nothing downloads.

Step 3: find the model in the picker

Open the chat tab and click the model picker. Kimi K3 sits among the other 37 names every plan and every pack carries, GLM 5.3, Qwen 3.8 Max, DeepSeek V4 Flash 0731 and GLM 5.3 Flash among them.

One thing to know before your first prompt: K3 reasons by default, and the Think button switches that off. On this model the switch is a spending decision rather than a style preference, which the next steps get into.

Step 4: use the context window on purpose

A million tokens of context is the feature most people underuse. It is not there so you can paste more; it is there so you can stop splitting work into chunks that lose the thread between them.

Two things it changes in practice. You can drop a whole repository subtree or a long contract in and ask questions across all of it, without a retrieval step in between deciding what the model gets to see. And you can let an agent run long enough to actually finish, instead of watching it forget the plan it wrote four tool calls ago.

The cost side of that is real and it comes up again in step 6: every turn in a long conversation resends the whole history, and input tokens are billed.

Step 5: paste screenshots

K3 accepts image input, and in a debugging loop that is the fastest thing the Studio does. A failing test output, a rendering bug, a chart with a number you cannot place: screenshot it into the composer and ask.

Not every model in the picker takes an image. If you attach one and the button is not there, you are on a text-only model, which is a property of that id rather than a setting you missed.

Step 6: turn thinking off when you do not need it

K3 reasons before it answers unless you tell it not to. That reasoning is output tokens, and output tokens on this model are the most expensive in our catalog.

Concretely: one million output tokens on K3 draws 1,425,000 credits, so a €5 pack of 165,000 credits is roughly 115,000 output tokens if you spend all of it here. A €19 month of 900,000 credits is roughly 630,000. Those are floors rather than forecasts, because input tokens come out of the same wallet.

The habit that pays: leave thinking on for the passes where the answer has to be worked out, and off for reformatting, extraction, summarising and anything where you already know the shape of the answer. On a short question the reasoning pass can be most of what you paid for.

The other habit: do not run a long agent loop on K3 by default. Most of what an agent generates is scaffolding nobody reads, and a cheaper model in the same picker produces the same scaffolding. Reach for K3 on the passes where the output decides something.

Step 7: call it from your own code

Go to Settings, then Cloud API keys, and create a key. It starts with lu_ and it is shown once, so paste it into your password manager before you close the dialog. The key spends your credits and cannot read or change your account.

The base URL is https://lu-labs.ai/api/inference/v1 and the shape is OpenAI chat completions, so anything that speaks that format works: Aider, LibreChat, your own script, or plain curl.

curl https://lu-labs.ai/api/inference/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"moonshotai/Kimi-K3",
       "messages":[{"role":"user","content":"Summarise this in one sentence: ..."}]}'

The model id is the catalog id, spelled exactly as above. Tool definitions go in the standard tools parameter; K3 calls them natively, so there is nothing special to configure.

Troubleshooting

The model is not in my picker. Check which account you are signed into. Any paid plan and any paid credit pack carries the whole catalog, K3 included. If the picker is short, you are probably in the demo Studio, which needs no account and does not carry the full list.

My API key returns 401. The key is shown once at creation. If you did not copy it, you cannot recover it, only revoke it and make a new one. Also check you sent it as Authorization: Bearer lu_... and not as a query parameter.

The reply starts with a pause. Cloud runs on shared hardware. At busy hours a request can sit in a queue for a few seconds before it starts streaming. That is queueing, not a stall, and it resolves on its own.

Credits went faster than I expected. Two usual causes. Thinking was on for everything, or a long conversation was resending a large history on every turn. Start a new chat when the topic changes; it is the cheapest habit on this list.

An image attachment is rejected. You switched to a text-only model. Check the picker before you attach.

Where to go next

If you are weighing K3 against the cheaper reasoners, Kimi K3 vs DeepSeek V4 puts them side by side with the prices. And if you want the model background rather than the how-to, we wrote that up when it launched: Kimi K3 explained.