How is Kimi K3 different from Kimi K2.7 Code?

K2.7 Code, released in June 2026, is a coding-focused model built on K2.6: 1T total and 32B active parameters, a 256K context, max_tokens defaulting to 32768, no reasoning_effort, and tool_choice limited to auto and none. K3, released July 16, moves to a new 2.8T architecture with a 1M context, up to 1048576 output tokens and adjustable reasoning, and widens the focus from coding to knowledge work and deep reasoning.

Why does sending temperature return an error?

Moonshot fixes K3 at temperature 1.0, top_p 0.95, n 1 and both penalties at 0; any other value returns an error. Its docs recommend leaving these parameters out of the request entirely.

Can I change reasoning_effort mid-conversation?

You can, but Moonshot warns that switching levels invalidates existing prefix-cache hits. Pick a level before the conversation starts; if reasoning takes too long, Moonshot suggests low.

Kimi K3

Save 30%

moonshot/kimi-k3

Kimi K3 — Moonshot AI's flagship open-weight reasoning model with a 1M-token context.

Get an API key View code sample

Live model playground

Try Kimi K3 here

Tune the request, see the current price ceiling, and inspect the real response without leaving this model page.

streamchat

1Configure input

Prompt

160 / 20,000

Add images

Maximum estimated cost—

2Live resultExample

Kimi K3

译文：

请在请求头中发送您的 API 密钥，并为每个请求设置合理的超时时间。

说明：

"请求头" 是简体中文技术文档中的标准译法（对应英文 "request header"）。
"携带" 可替换 "发送"，在 API 文档中更地道，如："请在请求头中携带您的 API 密钥……"
"合理的超时时间" 对应 "a sensible timeout"，也可视语境译为 "适当的超时时长"。

Context window: 1M
Max output: 1M
Input modalities: TextImage
Output modalities: Text
Capabilities: Vision · Reasoning · Tool calling · Coding · Prompt cache · Paid cache writes
Upstream release: 2026-07-16
RouteMux name: moonshot/kimi-k3

What is Kimi K3

Kimi K3 is Moonshot AI's flagship model, released on July 16, 2026. Moonshot calls it the first open 3T-class model. It has 2.8T parameters on a Stable LatentMoE design that activates 16 of 896 experts, with attention built on Kimi Delta Attention (hybrid linear attention) and Attention Residuals. Moonshot puts its overall scaling efficiency at roughly 2.5× Kimi K2 and positions it for long-horizon coding, end-to-end knowledge work and deep reasoning; its docs say that if you are unsure which model to use, start with kimi-k3. Moonshot reports 67.3 on DeepSWE.

The context window is 1M tokens, and max_completion_tokens defaults to 131072 and can be raised to 1048576. Vision is native, and on RouteMux it takes text and image input. Reasoning is always on and cannot be disabled; you set its depth with the top-level reasoning_effort field, which accepts low, high and max and defaults to max. Moonshot's parameter reference adds a few rules: temperature 1.0, top_p 0.95, n 1 and both penalties at 0 are fixed, other values return an error, so leave them out; in multi-turn chats and tool calls, send the full assistant message back unchanged, including reasoning_content; and tool_choice accepts auto, none and required. It is callable over OpenAI Chat Completions, OpenAI Responses and Anthropic Messages.

Against Kimi K2.7 Code: K2.7 Code, released in June, is a coding-focused model built on K2.6, with 1T total and 32B active parameters, a 256K context, no reasoning_effort support, and no required option for tool_choice. K3 is larger, has a far bigger context and output ceiling, and lets you tune reasoning depth.

On RouteMux, Kimi K3 costs $2.1 per 1M input tokens (official $3) and $10.5 per 1M output tokens (official $15), with cached input at $0.21 per 1M, 30% below the official price.

Before you choose

How is Kimi K3 different from Kimi K2.7 Code?: K2.7 Code, released in June 2026, is a coding-focused model built on K2.6: 1T total and 32B active parameters, a 256K context, max_tokens defaulting to 32768, no reasoning_effort, and tool_choice limited to auto and none. K3, released July 16, moves to a new 2.8T architecture with a 1M context, up to 1048576 output tokens and adjustable reasoning, and widens the focus from coding to knowledge work and deep reasoning.
Why does sending temperature return an error?: Moonshot fixes K3 at temperature 1.0, top_p 0.95, n 1 and both penalties at 0; any other value returns an error. Its docs recommend leaving these parameters out of the request entirely.
Can I change reasoning_effort mid-conversation?: You can, but Moonshot warns that switching levels invalidates existing prefix-cache hits. Pick a level before the conversation starts; if reasoning takes too long, Moonshot suggests low.

Kimi K3 API pricing

Prices verified 2026-10-12

InputOfficial $3 / M tokens

$2.1/ M tokensSave 30%

OutputOfficial $15 / M tokens

$10.5/ M tokensSave 30%

Cache read $0.21/M · Cache write $2.1/M

Estimate cost

Input tokensOutput tokens

Estimated cost$0.0315

Key capabilities

2.8T open weights: Stable LatentMoE with 16 of 896 experts active at a time; the full weights are public.
1M context, up to 1M output: max_completion_tokens defaults to 131072 and goes up to 1048576.
Three effort levels: Reasoning is always on; reasoning_effort is low, high or max, defaulting to max.
Native vision: One model understands text, images and video; on RouteMux it accepts images.
Three protocols: OpenAI Chat, OpenAI Responses, and Anthropic Messages all route to this model.

Common use cases

Coding agents: Moonshot recommends it for programming-agent work and positions it as its flagship for long-horizon coding.
End-to-end knowledge work: Knowledge work and deep reasoning are its other stated focus, shown through interactive research reports.
Long autonomous runs: Moonshot has shown it optimizing kernels, writing a GPU compiler and completing a 48-hour autonomous chip design.

Limited time10% extra Credits on your first top-up — up to 200 bonus Credits

Applies to your first top-up within 72 hours of signing up. Minimum $20, once per account.

Interface and protocols

OpenAI Chat CompletionsStreamingOpenAI ResponsesStreamingAnthropic MessagesStreaming

Use any of these IDs to call this model via the API.

RouteMux namemoonshot/kimi-k3

Replace the ROUTEMUX_KEY placeholder with your API key. Create one →

from openai import OpenAI

client = OpenAI(
    base_url="https://api.routemux.com/v1",
    api_key="<ROUTEMUX_KEY>",
)

completion = client.chat.completions.create(
    model="moonshot/kimi-k3",
    messages=[{"role": "user", "content": "What is the meaning of life?"}],
)
print(completion.choices[0].message.content)

base_url: https://api.routemux.com/v1

Performance (last 30 days)

67 requests · last 30 days

Total usage876.5K

P50 first-token latency8.10s

Avg throughput46 tok/s

Success rate100.0%

Uptime · 24htested via openai_chat

100.00%

Success rate · 60d · from real requests

Kimi K3 FAQ

›What is Kimi K3?

Kimi K3 — Moonshot AI's flagship open-weight reasoning model with a 1M-token context.

›How large is the context window of Kimi K3?

Kimi K3 supports a context window of up to 1M tokens, with up to 1M output tokens per request.

›How much does the Kimi K3 API cost?

Through RouteMux, Kimi K3 costs $2.1 for input and $10.5 for output (USD per 1M tokens), below the official list price.

›How do I call Kimi K3 via API?

Kimi K3 is OpenAI-compatible: point your base URL at RouteMux and set the model field to moonshot/kimi-k3 — no code changes needed.

›Which input and output modalities does Kimi K3 support?

Kimi K3 accepts text, image as input and produces text as output.

›When was Kimi K3 released?

Kimi K3 was released on July 16, 2026.

Sources

Reviewed by the RouteMux team on 2026-10-24

View as Markdown