Kimi K3

Available

A Moonshot AI model with a 1M-token context window, image input and tool calling. On the Kimi Open Platform API since July 16, 2026.

Access details: Kimi Open Platform API

The model ID to use in API requests.

kimi-k3

Open weights on Hugging Face: moonshotai/Kimi-K3

Updated Oct 1, 2026

Sources & notes

Capabilities & limits

Context window: 1M
Maximum output: 1M
Image input: Supported
Tool calling: Supported
Structured output: Supported

Pricing

USD / 1M tokens

Input: $3.00
Output: $15.00

Standard rates per million tokens. Discounts and long-context rates, where offered, are on Moonshot AI's pricing page.

Full rate card 5 rates

All prices in USD.

Input tokens (cache miss)

$3.00 per 1M tokens

Cached input tokens

$0.30 per 1M tokens

cache hit

5 minute cache writes

$3.00 per 1M tokens

context cache
5 minute TTL

1 hour cache writes

$6.00 per 1M tokens

context cache
1 hour TTL

Output tokens

$15.00 per 1M tokens

Prices exclude applicable taxes. The page lists a single K3 context tier (1,048,576 tokens). Cache hits refresh the entry's lifetime with no extra cache write charge.

Price history

Oct 1, 2026

Input $3.00 per million tokens

Oct 1, 2026

Output $15.00 per million tokens

Technical details

Parameters: 2.8T
Active parameters: 104B
Architecture: Kimi Delta Attention hybrid linear attention with Attention Residuals and a Mixture of Experts (16 of 896 experts active)
Training / knowledge cutoff: Not published
Input modalities: Text, image, video
Output modalities: Text

From Moonshot AI

Official announcement ↗

How they describe it

Kimi's most capable model to date

What they say is new

It is the world's first open-source model in the 3-trillion-parameter class

Release & lifecycle

First available: Jul 16, 2026
available: As of Oct 1, 2026