Moonshot AI: Kimi K3

Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work

Model lab: Moonshot AI
Model family: kimi-k3

Modalitiestext, image, video → text

Context window1,048,576

Maximum output131,072

Input · Quantization fp4USD 0.96 / 1M tokens

Output · Quantization fp4USD 7.2 / 1M tokens

Try now

Quickstart

Use your Moduck API Key, set the API base URL to https://api.moduck.ai/v1, and set the model parameter to moduck/moonshot-ai/kimi-k3.

import os
import requests

base_url = "https://api.moduck.ai/v1"
api_key = os.environ["MODUCK_API_KEY"]
model_id = "moduck/moonshot-ai/kimi-k3"

response = requests.post(
    f"{base_url}/chat/completions",
    headers={
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    },
    json={
        "model": model_id,
        "messages": [{"role": "user", "content": "Hello!"}],
        "stream": False,
    },
    timeout=60,
)
response.raise_for_status()
print(response.json())

Use the familiar OpenAI-style chat request format to make model integration easier.

Kimi K3 API Pricing

Type	Price information
Input · Quantization fp4	USD 0.96 / 1M tokens
Output · Quantization fp4	USD 7.2 / 1M tokens
Cached input · Quantization fp4	USD 0.32 / 1M tokens

Benchmarks

Max

Intelligence Index · v4.343.6

Coding Index76.2

Agentic Index50

Benchmark	Score	Measurement
DeepSWE	67.5	resolve rate · Variant: max effort · Harness: Kimi Code · Version: 1.1 · Date: 2026-07-16
Terminal-Bench	88.3	accuracy · Variant: max effort · Harness: Kimi Code · Version: 2.1 · Date: 2026-07-16
FrontierSWE	81.2	dominance score · Variant: max effort · Harness: Kimi Code · Date: 2026-07-16
Program Bench	77.8	score · Variant: max effort · Harness: Kimi Code · Date: 2026-07-16
SWE Marathon	42	resolve rate · Variant: max effort · Harness: Claude Code · Version: 1.1 · Date: 2026-07-16
GDPval-AA	1,668	Elo · Variant: max effort · Version: v2 · Date: 2026-07-16
AA-Briefcase	1,548	Elo · Variant: max effort · Date: 2026-07-16
AutomationBench	30.8	success rate · Dataset: 600-task public subset · Variant: max effort · Date: 2026-07-16
JobBench	52.9	score · Variant: max effort · Date: 2026-07-16
SpreadsheetBench	34.8	score · Variant: max effort · Harness: Claude Code · Version: 2 · Date: 2026-07-16
BrowseComp	91.2	accuracy · Variant: max effort, context compaction · Date: 2026-07-16
CharXiv Reasoning	91.3	accuracy · Variant: max effort, with tools · Date: 2026-07-16
ZeroBench	41	pass@5 · Variant: max effort, with tools · Date: 2026-07-16

Performance

Measured on the Moduck gateway across live traffic, hour by hour over the last 24 hours. Time to first token is how long before anything appears; throughput is how fast the answer is written after that; end-to-end latency is the whole round trip. Figures are medians unless the card says otherwise.

Artificial Analysis measurement · Max

Output speed42.16 tokens/s

Time to first token3.63 s

Time to first answer token—

End-to-end response time—

Measurement variant	Output speed	Time to first token	Time to first answer token	End-to-end response time
Max	42.16 tokens/s	3.63 s	—	—

Performance trends

Throughput42.16 tok/sCurrent

Time to first token3.63 sCurrent

End-to-end latency7.9 sCurrent

Throughput

Current42.16 tok/s

Window average42.4 tok/s

Time to first token

Current3.63 s

Window average4.2 s

End-to-end latency

Current7.9 s

Window average7.5 s

Tail latency (P99)

Current32.4 s

Window average17.4 s

Prompt cache hit rate

Current19.3%

Window average24.0%

Error rate

Current0.0%

Window average0.2%

Availability

Uptime is the share of time at least one provider serving this model responded. Availability is stricter: the share of requests actually served. A provider that answers but fails part of the traffic lowers availability without lowering uptime. Both cover the last three days.

Uptime100%At least one provider responded

Availability100%Requests actually served

Error rate0%Share of failed requests

Served requests, hour by hour

24 hours agoNow

Kimi K3 Compare models

Model	Input price	Output price	Context window	Capabilities
Kimi K3This page	USD 0.96 / 1M tokens · Quantization fp4	USD 7.2 / 1M tokens · Quantization fp4	1,048,576	Text output · Reasoning · Tool calling
Kimi K2.7 Code	USD 0.44 / 1M tokens	USD 1.8 / 1M tokens	262,144	Text output · Reasoning · Tool calling
Kimi K2.7 Code Highspeed	USD 1.52 / 1M tokens	USD 6.4 / 1M tokens	262,144	Text output · Reasoning · Tool calling
Kimi K2.6	USD 0.4 / 1M tokens	USD 1.592 / 1M tokens	262,144	Text output · Reasoning · Tool calling
Kimi K2.5	USD 0.324 / 1M tokens	USD 1.584 / 1M tokens	262,144	Text output · Reasoning · Tool calling

Model specifications

Input: text, image, video
Output: text
Context window: 1,048,576
Maximum output: 131,072
Release date: 2026-07-16
Attachments: Yes
Reasoning: Yes
Tool calling: Yes
Structured output: Yes
Temperature setting: No
Open weights: Yes

Frequently asked questions

What is Kimi K3 for?

Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work

What inputs and outputs does Kimi K3 support?

Recorded inputs: text, images, video; outputs: text.

How is Kimi K3 priced?

The pricing table shows the current recorded Moduck price. For example: Input · Quantization fp4 USD 0.96 / 1M tokens; Output · Quantization fp4 USD 7.2 / 1M tokens. See the pricing table for all units, variants, and conditions.

What specifications are recorded for Kimi K3?

Current model records list Context window 1,048,576 tokens; Maximum output 131,072 tokens.

What model parameter does the Quickstart use?

The example uses moduck/moonshot-ai/kimi-k3. This is a cataloged Moduck public or reserved ID; check your account’s available models before calling it.

Can I call this model through the Moduck API now?

This page has no verified Moduck API call for this model. A Quickstart example or reserved ID does not establish access; check the models available to your account.