Hosted in Australia
Kimi K3 API,
hosted in Australia.
Moonshot AI’s Kimi K3 behind one OpenAI-compatible API, served from dedicated GPUs in Australia. Same per-token price as the model vendor, billed per token.
Pricing
Kimi K3 API pricing
Billed per token, in USD per 1M tokens.
These are the same input and output prices Moonshot AI lists for Kimi K3 on its own API (as of October 2026). See pricing for every model, or compare both models on the Model API page.
Model facts
Kimi K3 at a glance
Model details from Moonshot AI’s own model card and API docs, as of October 2026.
| Detail | Kimi K3 |
|---|---|
| Model | Kimi K3 |
| Maker | Moonshot AI |
| Architecture | Mixture-of-Experts. 2.8T total parameters, 104B activated per token. |
| Context window | 1,048,576 tokens (1M). Model maker’s figure |
| Max output | Up to 1,048,576 tokens per request on Moonshot AI’s own API, with a default of 131,072. Model maker’s figure |
| What it’s good for | Frontier reasoning and agentic workloads, including long-horizon coding and multi-step research. |
Sources: Moonshot AI, Kimi K3 model card and Kimi K3 quickstart, checked 23 October 2026. Where a row says “model maker’s figure”, the limit on Xinference may differ. Kimi and the Kimi logo are trademarks of Moonshot AI, shown only to identify the model. Their use does not imply endorsement.
Quickstart
Call Kimi K3 from Python or curl
Your base URL and key come with your API access. Set them as environment variables, then use the OpenAI SDK or curl.
# pip install openai
import os
from openai import OpenAI
client = OpenAI(
base_url=os.environ["XINFERENCE_BASE_URL"],
api_key=os.environ["XINFERENCE_API_KEY"],
)
response = client.chat.completions.create(
model=os.environ["XINFERENCE_MODEL"],
messages=[
{"role": "user", "content": "Summarise this clause."},
],
)
print(response.choices[0].message.content)curl $XINFERENCE_BASE_URL/chat/completions \
-H "Authorization: Bearer $XINFERENCE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "'"$XINFERENCE_MODEL"'",
"stream": true,
"messages": [{"role": "user", "content": "Hello"}]
}'
# tokens stream back as server-sent events.Use the model name listed with your API access. Moonshot AI’s docs note that Kimi K3 always reasons first, and that multi-turn and tool-calling requests should pass the full assistant message back unchanged. See the Kimi K3 quickstart.
Why Xinference
Why run Kimi K3 on Xinference
One OpenAI-compatible API
Kimi K3 and DeepSeek V4.1 Flash sit behind the same endpoint. Code written for the OpenAI SDK needs a new base URL and key, not a rewrite. What an OpenAI-compatible API covers.
Hosted in Australia on dedicated GPUs
Kimi K3 is served from dedicated GPUs in Australia.
Same per-token price as the model vendor
$3.00 in and $15.00 out per 1M tokens, the same as Moonshot AI’s list price for its own API.
Billed per token
You pay for the input and output tokens you use, in USD per 1M tokens. See pricing.
No training on your data
We don’t use your prompts or outputs to train or fine-tune models unless you agree in writing. No prompt or response content kept by default.
Questions
How much does the Kimi K3 API cost?
On Xinference, Kimi K3 costs $3.00 per 1M input tokens and $15.00 per 1M output tokens, in USD. That is the same per-token price Moonshot AI lists for its own API as of October 2026. You are billed per token. See pricing for every model on the Model API.
Is there a Kimi K3 API hosted in Australia?
Yes. Xinference serves Kimi K3 from dedicated GPUs in Australia, through an OpenAI-compatible API. You call it with the OpenAI SDK or curl, using the base URL and API key that come with your API access. DeepSeek V4.1 Flash is available on the same API.
Is the Kimi K3 API OpenAI-compatible?
Yes. The Model API accepts OpenAI-format requests on /v1/chat/completions, so existing OpenAI SDK clients work with a new base URL and key. Streaming, function calling and structured output (JSON mode) go through the same API. What an OpenAI-compatible API covers.
What is the Kimi K3 context window?
Moonshot AI lists Kimi K3 with a context window of 1,048,576 tokens, about 1M, as of October 2026. That is the model maker’s figure, and the limit on Xinference may differ.
Do you train on my data?
No. We don’t use your prompts or outputs to train or fine-tune models unless you agree in writing. No prompt or response content kept by default. This applies to Kimi K3 and DeepSeek V4.1 Flash alike. The Terms of Service and Privacy Policy set out the details.
Can I use other models on the same API?
Yes. The Model API also serves DeepSeek V4.1 Flash at $0.30 in and $1.20 out per 1M tokens, suited to long-context work and RAG. You switch models by changing the model name. Don’t see the model you need? Talk to us about adding it.
Last updated