Make your first Kimi K3 API call
Model ID kimi-k3, $3.00 in / $15.00 out per 1M tokens, a 1,048,576-token window. The exact code that works is below — copy it, add a key, and your first call runs.
{
"model": "kimi-k3",
"choices": [{
"message": {
"content": "OK",
"reasoning_content": "We need answer to user…"
},
"finish_reason": "stop"
}],
"usage": {
"prompt_tokens": 116,
"completion_tokens": 53,
"completion_tokens_details":
{ "reasoning_tokens": 37 }, // billed as output
"prompt_tokens_details":
{ "cached_tokens": 116 } // 10× cheaper
}
}- Model ID
- kimi-k3
- Input
- $3.00 /1M
- Cached input
- $0.30 /1M
- Output
- $15.00 /1M
- Context
- 1,048,576
- Architecture
- 2.8T · KDA hybrid linear attention · 16 of 896 experts active · official quickstart, read 2026-07-20
- Open weights
- by 2026-07-27 · official quickstart, read 2026-07-20
- Success rate · 24 h
- live →
Kimi K3 is Moonshot AI's flagship model, launched July 2026. The API is live now: model ID kimi-k3, a 1,048,576-token context window, $3.00 per 1M input tokens ($0.30 cached) and $15.00 per 1M output. It speaks the OpenAI format, and EvoLink carries it at the same rates — no markup, one base_url change.
Under the hood: 2.8 trillion total parameters in a mixture-of-experts design (16 of 896 experts active per token), built on Kimi Delta Attention, with native visual understanding. The open weights are officially due by July 27, 2026 — we're tracking the drop.
Prices from the official page, 2026-07-20; architecture from the official quickstart, 2026-07-18. Everything marked measured, we ran ourselves — how.
We hit the potholes so you don't have to
We ran kimi-k3 within hours of its July 2026 launch and mapped the sharp edges. Get three settings right and your integration works the first time, at the price you expected.
Input caching is automatic
Repeat your prefix and input drops from $3.00 to $0.30 per 1M — no parameter, no code change. We measured exactly how it triggers, so you can keep it hot.
Keep the cache hot Before you shipOne setting protects your bill
Leave max_completion_tokens at its default and empty-response billing can't happen to you. If you must cap it, we measured where the floor is.
Output includes thinking — plan for it
K3 reasons on every request; that's the model's power, and it bills as output. Budget a few times your answer length and your numbers come out right.
See the breakdownChange one line and it runs
K3 speaks the OpenAI format. Point base_url at a gateway that carries it, set the model, done.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_EVOLINK_API_KEY",
base_url="https://direct.evolink.ai/v1", # <-- the one line you change
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Explain prompt caching in two sentences."}],
)
print(response.choices[0].message.content)Leave max_completion_tokens at its 131,072 default unless you have read why capping it backfires. Full walkthrough, auth, and streaming: the API guide.
Go deeper
One page per question. Every number on them is either quoted from the official docs with a date, or measured by us.
Try Kimi K3 free
The chat route and the API route with 10 free signup credits — and how to make them go further.
Start free ImplementKimi K3 API guide
Auth, the request that works, streaming, what a real response body looks like, and how to tell a success from a silent failure.
Start calling Choose a routeEvoLink vs OpenRouter
Both carry K3 at the same price. Same suite through both, one day: field names, cache, reliability, latency.
See the measured comparison CheckKimi K3 status
Success rate and latency, charted live over time.
See the charts BudgetPricing
The official rates with a date stamp — and the reasoning-token overhead that makes the output rate the one that matters.
See real costs InteractiveCost calculator
Your requests, your prompt structure, your hit rate — a monthly bill at the official rates, computed in your browser.
Run your numbers WatchOpen weights: July 27
Officially committed, not yet delivered. The status tracker, the 2.8T hardware math, and a ready-for-the-drop checklist.
Track the drop ScaleThe 1M context window
Half a million tokens in one call, measured — and how a warm cache makes repeat runs cost a tenth. When the window pays, it really pays.
See the numbers EvaluateLatency
First token in ~3s on trivial prompts (2.8s on 07-16, 3.35s on 07-18); realistic prompts: 21.4s median to first content token — and the streaming setup that hides all of it.
See the timings DebugErrors & failure modes
Real response bodies from deliberately broken calls — including two bad parameters that return HTTP 200 instead of erroring.
Fix itKimi K3 FAQ
Short answers — each with a dated official source or one of our measurements behind it.
What is Kimi K3?
Kimi K3 is Moonshot AI's flagship large language model, launched July 2026: 2.8 trillion total parameters in a mixture-of-experts design (16 of 896 experts active per token), built on Kimi Delta Attention, with native visual understanding and a 1,048,576-token context window. Per the official quickstart, read 2026-07-18.
Is Kimi K3 open source?
Not yet. The weights are officially committed for release by July 27, 2026, but as of 2026-07-18 they are not published and no license has been announced. We're tracking the release and will update within hours of the drop.
How much does the Kimi K3 API cost?
$3.00 per 1M input tokens ($0.30 on a cache hit) and $15.00 per 1M output tokens — and reasoning tokens bill as output, which in our runs was 43–77% of it. Official pricing page, read 2026-07-20. The full breakdown.
How do I access Kimi K3?
Three ways: chat at kimi.com; the official API on Moonshot's platform; or an OpenAI-compatible gateway like EvoLink, which carries it at the same rates — for existing OpenAI-format code it's a one-line base_url change. First call in five minutes.
How big is the Kimi K3 context window?
1,048,576 tokens — a full binary million. max_completion_tokens defaults to 131,072; capping it low is how you get empty responses billed in full. What using the window actually costs.
Does Kimi K3 handle images?
The official quickstart lists native visual understanding (read 2026-07-18). We haven't measured vision workloads yet — everything we have measured is documented in our method.
Get a key and make the call
EvoLink carries kimi-k3 on an OpenAI-compatible endpoint — one key reaches GPT, Claude, Gemini, and dozens of the world's mainstream models. 10 free credits, sign up from anywhere.