Kimi K3 API

Kimi K3 reads text, images and video, and works through long coding and agent runs, from $2.250 per million input tokens.

Get API key Read the docs

1M token context
2.8T total parameters
FP8 weights served
Kimi K3 license

from openai import OpenAI

client = OpenAI(
    base_url="https://api.modelsheep.com",
    api_key=API_KEY_MODELSHEEP,
)

response = client.chat.completions.create(
    model="moonshotai/Kimi-K3",
    messages=[
        {
            "role": "user",
            "content": "Why do sheep count people?"
        }
    ],
)

print(response.choices[0].message.content)

How to use the Kimi K3 API

1
Create a ModelSheep account and get an API key from the dashboard.
2
Point the OpenAI SDK at the base url https://api.modelsheep.com and set the model to moonshotai/Kimi-K3.
3
Call /chat/completions with your messages, and the reply returns the answer plus a reasoning trace that reasoning_effort controls.

What Kimi K3 is

Kimi K3 is an open-weight model from Moonshot AI. It uses a mixture of experts design that activates 16 of 896 experts for each token, so one request uses a small share of its 2.8T weights. Moonshot built it for long coding sessions, agentic knowledge work and reasoning, and the same model reads images and video beside text. Thinking is always on, so every reply carries a reasoning trace.

What it does well

Runs long coding sessions across large repositories and terminal tools.
Answers research questions requiring many browsing steps.
Reads a document page or video frame and answers questions on it.

What it is not for

Short, cheap calls. You cannot switch thinking off. Every request pays for
A large hosted service without a contract. The license makes a model as a

Calling it

Post to /chat/completions with the model string moonshotai/Kimi-K3. Set reasoning_effort to low, high, or max. The default is max. It spends the most tokens. Callers miss one step. In a chat with more than one turn, or in a tool loop, send the whole assistant message back in messages. Keep reasoning_content and tool_calls intact. Moonshot trained the model in preserved thinking history mode. It requires that history.

moonshotai/Kimi-K3 /chat/completions

SDKs and languages

Kimi K3 answers on an HTTP endpoint, so any language that can post JSON reaches it. These are the ones the box above writes out.

OpenAI SDK: PythonTypeScriptGoJavaC#PHPRust
Anthropic SDK: PythonTypeScriptGoJavaC#PHPRust
HTTP: cURLPythonTypeScriptGoRustJavaC#PHP

Specifications

Author	Moonshot AI
Category	language
Accepts	text, image, video
Returns	text
Total parameters	2.8T
Active parameters	dense, all of them
Score	43.8 on Artificial Analysis intelligence index
Languages	not published
Precision served	FP8
Context	1M tokens
Vector	n/a
License	Kimi K3

Price

Input

$2.250

per million tokens

Output

$11.475

per million tokens

Against OpenRouter

10%

below the cheapest of 13 providers

The cheapest OpenRouter price for this model was $2.500 in and $12.750 out, with a median of $3.000 and $15.000 across the comparable providers. Read on 2026-09-17.

License

Moonshot publishes the weights under the Kimi K3 license. It permits commercial use, but a model as a service business above a revenue threshold must sign a separate agreement with Moonshot AI, and a product above a user or revenue threshold must show the model name in its interface. This is not a standard open license. Read the license file before you ship on it.

Similar models

GLM 5.3zai-org/GLM-5.3

Parameters: 753.3B
Context: 1.3M
Input per 1M: $0.915
Output per 1M: $2.875
License: GLM 5.3

GLM 5.3 Flashzai-org/GLM-5.3-Flash

Parameters: 321.3B, 18B
Context: 1.3M
Input per 1M: $0.081
Output per 1M: $0.270
License: MIT

Qwen3.8 2.4T A95BQwen/Qwen3.8-2.4T-A95B

Parameters: 2.4T, 95B
Context: 1M
Input per 1M: $1.800
Output per 1M: $5.400
License: Qwen3.8 Max

Start using Kimi K3

Point the OpenAI SDK at api.modelsheep.com. Your prompts stay in Germany, and nothing is retained.

Get API key Read the quickstart