Kimi K3 API: Moonshot's 2.8T, 1M-Context Model via an OpenAI-Compatible Endpoint

#llm #ai #api #webdev

Disclosure: I run NovAI, one of the API gateways that serves Kimi K3. Everything below (specs, pricing, code) is verifiable — specs are from Moonshot's announcement, and you can check pricing yourself.

Moonshot AI released Kimi K3 on 2026-07-16: a 2.8-trillion-parameter MoE, natively multimodal (text + vision), with a 1M-token context window, and open weights announced for July 27. It's the most interesting Chinese model drop of the summer — here's how to actually call it over an API today.

TL;DR

If you already use the OpenAI SDK, it's a two-line change:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://aiapi-pro.com/v1",  # one gateway serving K3
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain mixture-of-experts routing in 3 bullet points."}],
)
print(resp.choices[0].message.content)

Streaming works exactly as you'd expect:

for chunk in client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Write a haiku about long context."}],
    stream=True,
):
    print(chunk.choices[0].delta.content or "", end="", flush=True)

And because K3 is natively multimodal, vision input uses the standard image_url content type:

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What's in this chart?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
        ],
    }],
)

Why K3 matters

Per Moonshot's announcement:

2.8T total parameters (MoE) — the largest open-weights model announced to date
1M-token context — a full codebase or a stack of PDFs in one prompt
Native multimodality — vision is built in, not bolted on
Open weights — self-hostable once weights drop (announced July 27)

Pricing (honest numbers)

On NovAI, K3 is $2.90 / $14.50 per million tokens (input / output) for the 1M-context tier. For comparison, OpenRouter lists K3 at roughly $3.00 / $15.00 — so we're about 3% cheaper. That's parity, not a price war; the point of a gateway is one API surface across every major Chinese lab (DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao, Hunyuan) plus image and video generation.

Check current numbers yourself: https://aiapi-pro.com/pricing