Kimi K3 API

MODEL NAME

kimi-k3

BASE URL

https://api.moonshot.ai/v1

WORKS WITH

The OpenAI SDK. An Anthropic-style address is also offered

CONTEXT

1,048,576 tokens

OUTPUT LIMIT

131,072 tokens by default (max_completion_tokens)

THINKING

Always on. reasoning_effort: low, high or max

INPUT

Text, images and video

PRICE PER MILLION TOKENS

$3.00 in · $0.30 cached · $15.00 out

From Moonshot's Kimi K3 quickstart and price page. Prices exclude tax.

1. Sign in to Moonshot's API platform at platform.kimi.ai and add a small balance. The minimum is $1.
2. Create an API key in the console. Copy it straight away.
3. Save it as an environment variable named MOONSHOT_API_KEY. Do not paste it into your code.
4. For Python, install the OpenAI SDK: pip install --upgrade "openai>=1.0". For Node, run npm install openai.

Need more detail on keys and accounts? See Kimi API setup.

All three examples ask Kimi K3 a simple question and print the answer. They use low effort and a small output limit to keep the test cheap.

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -d '{
    "model": "kimi-k3",
    "reasoning_effort": "low",
    "max_completion_tokens": 4096,
    "messages": [
      {"role": "user", "content": "Explain what an API is in two sentences."}
    ]
  }'

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

response = client.chat.completions.create(
    model="kimi-k3",
    reasoning_effort="low",
    max_completion_tokens=4096,
    messages=[
        {"role": "user", "content": "Explain what an API is in two sentences."}
    ],
)

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

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.MOONSHOT_API_KEY,
  baseURL: "https://api.moonshot.ai/v1",
});

const response = await client.chat.completions.create({
  model: "kimi-k3",
  reasoning_effort: "low",
  max_completion_tokens: 4096,
  messages: [
    { role: "user", content: "Explain what an API is in two sentences." },
  ],
});

console.log(response.choices[0].message.content);

Written for this guide in the standard OpenAI format, using the address, model name and settings from Moonshot's quickstart. Notice what is missing: no temperature setting. An older OpenAI SDK may not accept reasoning_effort or max_completion_tokens, so update it first.

Kimi K3 always thinks before it answers, and you cannot switch that off. What you can set is how much effort it spends. More effort usually means better answers on hard problems, with a longer wait and a bigger bill, because thinking is charged as output.

The three values and the default are from Moonshot's reasoning effort guide. The advice on when to use each is ours. Moonshot's guide does not give figures for the difference, so measure it on your own tasks.

Moonshot fixes several settings for Kimi K3 and says to leave them out of requests.

On a small screen, swipe the table sideways to see every column.

Many libraries send a temperature by default. If yours does, look for a way to remove it. Need thinking switched off completely? Use kimi-k2.6 for those calls.

stream = client.chat.completions.create(
    model="kimi-k3",
    reasoning_effort="high",
    stream=True,
    messages=[{"role": "user", "content": "Plan a 3-step database migration."}],
)

for chunk in stream:
    if not chunk.choices:
        continue
    delta = chunk.choices[0].delta
    thinking = getattr(delta, "reasoning_content", None)
    if thinking:
        print(thinking, end="", flush=True)   # the model's reasoning
    if delta.content:
        print(delta.content, end="", flush=True)  # the answer

Continues from the Python example above. A teaching example written for this guide.

Tools let Kimi K3 do real work: search, read files, run code. Moonshot's guides set out a few rules that matter more for this model than for older ones.

For a full tool loop in Python, see the example on the Kimi K2.7 Code API page. The pattern is the same.

{
  "model": "kimi-k3",
  "prompt_cache_options": {"mode": "implicit", "ttl": "1h"},
  "messages": [
    {"role": "system", "content": "LONG FIXED INSTRUCTIONS AND REFERENCE TEXT"},
    {"role": "user", "content": "The question that changes each time"}
  ]
}

From Moonshot's context caching guide. Text is cached in blocks, and a part smaller than one block is charged at the normal price. The guide does not give the block size.

Enter a month's usage to see the bill, next to the same work on Kimi K2.6.

List prices before tax. It leaves out the one-off charge for saving to the cache and any web search calls. The Kimi K2.6 figure uses $0.95 in, $0.16 cached and $4.00 out. Kimi K3 always thinks before answering, so its output may be higher than Kimi K2.6 with thinking off.

Moonshot's batch price page lists only Kimi K2.7 Code and Kimi K2.6, so Kimi K3 does not appear to be offered at the cheaper batch rate. More: API pricing · rate limits and enterprise pricing · API vs subscription

On a small screen, swipe the table sideways to see every column.

From Moonshot's error code page and Kimi K3 guides. More help: Kimi not working?