Can I run Kimi K3 on my own PC?

Realistically, no. Kimi K3 is a 2.8 trillion parameter model. No consumer card and no single workstation holds it at any quant, and the hardware math is spelled out in our can-you-run-Kimi-K3-locally post. A hosted endpoint is the working answer.

Do I need a subscription to use Kimi K3?

No. Kimi K3 is on every LU Labs plan including the pay as you go wallet, so a 5 EUR credit pack with no subscription reaches it. Pack credits never expire.

What does Kimi K3 cost per token?

One credit is $0.00001 of wholesale cost. Kimi K3 draws 0.285 credits per input token and 1.425 per output token, so 285,000 credits per million tokens in and 1,425,000 per million out. A 5 EUR pack of 165,000 credits is roughly 115,000 output tokens on it.

What is the model id for Kimi K3 on the LU Labs API?

moonshotai/Kimi-K3. The base URL is https://lu-labs.ai/api/inference/v1 and the request shape is OpenAI chat completions, so an existing client usually needs only a changed base URL, key and model string.

Does Kimi K3 take images?

Yes. Image input works in the browser Studio and over the API as an image content part. It also holds a 1M token context and calls tools natively, so it runs in Agent and coding modes.

Can I turn the thinking off?

Yes. The think toggle in the Studio maps to the reasoning effort the API takes, and on this model switching it off genuinely stops the reasoning pass. On a short factual question we measured 34 completion tokens dropping to 8.

Where do I get an API key?

In your LU Labs account under Settings and then Cloud API keys. The key carries an lu_ prefix and is displayed once at creation. If you lose it, delete the key and create a new one.

Kimi K3 online: browser, desktop app and API key

We build a desktop app whose whole point is running models on your own machine, so this costs us something to write: Kimi K3 is not a model you run at home. It is 2.8 trillion parameters. There is no quant that fixes that, and the arithmetic is laid out in can you run Kimi K3 locally if you want to see it fail in detail.

So this page is about the other side of the house: the three ways to reach Kimi K3 on hardware you do not own. If you want to know what the model actually is first, that is Kimi K3 explained.

The short version

Kimi K3 sits on every LU Labs plan, including the pay as you go wallet. That means a 5 EUR credit pack with no subscription reaches it. It takes images as input, it has a think toggle, it holds a 1M token context, and it calls tools natively, so the Agent and the coding agent both run on it.

Route one, the browser

Open lu-labs.ai and sign in. Nothing to install, no driver, no GPU on your side.
Put credits on the account. The smallest pack is 5 EUR for 165,000 credits and needs no subscription. Prices sit on the pricing page.
Open the model picker above the chat input and choose Kimi K3.
Type. Attach a picture if you have one, the model reads image input. Leave the think toggle on for hard prompts and off for short ones, because the reasoning pass is output tokens and output tokens are the bill.

Route two, the desktop app

The free open source app for Windows and Linux has a Cloud switch that uses the same account. The point of that switch is that you do not run two workflows: a 27B on your own card for the private everyday work, one click for Kimi K3 when the prompt is bigger than the machine, one click back. Same chat window, same history, same key.

Install the app, sign in under the cloud settings, flip the switch, pick Kimi K3 in the same picker your local models live in. The comparison against the usual desktop runners is in LM Studio alternatives.

Route three, an API key

Sign in and go to Settings > Cloud API keys.
Create a key. It carries an lu_ prefix and is shown once, so copy it before the dialog closes. Lose it and you delete the key and make a new one.
Point your client at base URL https://lu-labs.ai/api/inference/v1.
Use the OpenAI chat completions shape with the model id moonshotai/Kimi-K3.

curl https://lu-labs.ai/api/inference/v1/chat/completions \
  -H "Authorization: Bearer $LU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/Kimi-K3",
    "messages": [
      {"role": "user", "content": "Rewrite this paragraph in a colder register."}
    ]
  }'

The same thing with the official OpenAI Python SDK, which needs no LU specific library:

from openai import OpenAI

client = OpenAI(
    api_key="lu_your_key_here",
    base_url="https://lu-labs.ai/api/inference/v1",
)

resp = client.chat.completions.create(
    model="moonshotai/Kimi-K3",
    messages=[{"role": "user", "content": "Draft three openings for a heist scene."}],
)
print(resp.choices[0].message.content)

Clients that take a base URL

Client	What to set
SillyTavern	API type Chat Completion, source Custom (OpenAI-compatible). Base URL and key as above, model id in the custom model field.
Aider	OpenAI compatible base URL and key in the environment, model id on the command line.
LibreChat	A custom endpoint with the base URL, the key, and the model id in the model list.
Anything else	If it takes a base URL, a key and a model string, it works. That is most of the tooling built in the last two years.

What it costs

Billing is in credits and 1 credit is $0.00001 of wholesale inference cost. Kimi K3 draws 0.285 credits per input token and 1.425 per output token, which is 285,000 credits per million tokens in and 1,425,000 per million out. It is the most expensive model in the catalog, and that is the honest reason it sits on the entry plan instead of behind an upgrade: the meter charges per token, so nobody is locked out, they just spend faster.

What you buy	Credits	Output tokens on Kimi K3
5 EUR pack, no subscription	165,000	about 115,000
10 EUR pack	350,000	about 245,000
25 EUR pack	900,000	about 630,000
Hosted plan, 19 EUR a month	900,000 a month	about 630,000

Input tokens draw from the same wallet, so a long chat history lands below those numbers. Pack credits never expire and the Hosted plan adds two character trainings a month, which packs do not include. If those figures make you wince, that is the correct reaction and the answer is a cheaper model for the volume work: DeepSeek in the cloud covers the other end of the price range.

Thinking and images

Kimi K3 reasons when you ask it to and stays quiet when you do not. The think toggle in the Studio maps to the reasoning effort the API takes, and switching it off really does stop the reasoning on this model, which is not true of every reasoner in the catalog. On a short factual question the difference we measured was 34 completion tokens down to 8.

Image input works. Attach a picture in the Studio, or send an image content part over the API the way you would to any OpenAI shaped endpoint.

What you give up

Shared hardware. Managed H100, A100 and B200 class GPUs, shared across users, so at a busy hour you can sit in a short queue. A dedicated box would not do that, and it would not cost 5 EUR either.
A curated catalog. 38 chat models, all of them on every plan and on a credit pack, but no bring your own checkpoint. If your workflow depends on one specific community fine tune, local is your route.
Somebody else's machine. Accounts and billing sit in the EU, prompts are not used for training and are never sold, and the hosted text models answer as they were trained with no moderation layer added on top. Image and video follow the Terms of Service and the prompt is checked before it renders. Content involving minors is refused everywhere, always.