Kimi K3 on Runpod

Moonshot AI's 2.8T-parameter open-weight model, ready to deploy in minutes. Call it through Runpod's OpenAI-compatible endpoint, or self-host it on multi-node GPUs.

Deploy on Runpod

View deployment guide

Model details

Developed byMoonshot AI

Model familyKimi

Use caseLong-horizon coding and agentic knowledge work

VersionK3

Size2.8T parameters (16 of 896 experts active)

Context1M tokens

Price$3.00 / 1M input tokens
$15.00 / 1M output tokens

LicenseKimi-K3 model license

Example usage

Kimi K3 is available now as a Runpod Public Endpoint. It speaks the OpenAI Chat Completions format, so point your existing client at Runpod and pass model kimi-k3. Streaming is supported.

Python (OpenAI SDK) cURL (direct HTTP)

Using the OpenAI SDK pointed at Runpod's OpenAI-compatible base URL.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.runpod.ai/v2/moonshot-kimi/openai/v1",
    api_key="RUNPOD_API_KEY",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "system", "content": "You are Kimi."},
        {"role": "user", "content": "Explain what makes Kimi K3 good at long-horizon coding."},
    ],
    stream=True,
)

for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="")

A direct HTTPS request to the OpenAI-compatible endpoint.

curl https://api.runpod.ai/v2/moonshot-kimi/openai/v1/chat/completions \
  -H "Authorization: Bearer $RUNPOD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {"role": "user", "content": "What is Runpod?"}
    ]
  }'

Example response

{
  "id": "chatcmpl-8f3a2b",
  "object": "chat.completion",
  "model": "kimi-k3",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Runpod is an AI developer cloud for building, training, and serving models on GPUs."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 12, "completion_tokens": 18, "total_tokens": 30 }
}

Request flow

How a Kimi K3 request flows on Runpod

Point your OpenAI client at the Runpod endpoint, Runpod routes the call to Kimi K3, and tokens stream back to your app.

Deploy your own Kimi K3 cluster

Kimi K3 is a 2.8T-parameter model with a total footprint around 1.56TB of VRAM, so it runs across multiple GPUs rather than a single card. Pick a supported deployment shape below and bring a day-0 serving recipe. Larger clusters work too, these are supported starting points, not hard limits.

Supported deployment shapes on Runpod

B200 2x8

GB200 4x4

H100 4x8

B300 1x8

H200 2x8

GB300 2x4

MI350X / MI355X 1x8

Day-0 serving recipes are available for vLLM and SGLang.

vLLM recipe SGLang recipe

# Day-0 vLLM image for Kimi K3
docker pull vllm/vllm-openai:kimi-k3

# Serve command: follow the day-0 recipe for the exact flags.
# [VLLM SERVE FLAGS PENDING] confirm from recipes.vllm.ai/moonshotai/Kimi-K3

SGLang reports 423 tokens per second for Kimi K3 on DSpark. This is SGLang's reported figure, not a Runpod-measured benchmark.

Related models

Kimi K2.6

Default Kimi endpoint for reasoning and chat.

$0.95 / 1M input tokens
$4.00 / 1M output tokens

Kimi K2.7 Code

Tuned for coding and agentic workflows.

$0.95 / 1M input tokens
$4.00 / 1M output tokens

Build what’s next.

Build, train, and scale AI workloads on Runpod with cloud GPUs, Serverless, and Clusters.

Get started