NezhaGateNezhaGate
简体中文English ✓Español한국어日本語
Log in Get API Key
Chat

Kimi K3

⧉
kimi-k3

Kimi K3 is the new flagship from Moonshot AI, strong at long-document understanding, multi-document analysis and Chinese writing, with excellent agent and tool use. It always thinks before it answers, with the reasoning in reasoning_content; the sampling temperature is fixed by the model, so temperature / top_p in the request are ignored. Tool calling, streaming and prompt caching are supported (a repeated long prefix settles at the cache rate). Fully OpenAI-compatible: set model to kimi-k3. Pay-as-you-go, failed calls never billed.

ChatLong contextAgentsTool callingOpenAI compatible
Get API Key → API Docs

Why this model

Moonshot's new flagship: 2.8T parameters, open weights, a context of about 1M tokens

Highlights

  • 2.8T parameters with open weights, one of the largest open-weight models to date
  • About 1M tokens of context (1,048,576): hand it a whole long document
  • Thinks by default; the reasoning comes back separately in reasoning_content, never mixed into the answer
  • Tool calling, streaming and prompt caching; a repeated long prefix bills at the cache rate
  • OpenAI-compatible: existing SDK code only changes base_url and model

Price vs. official list

NezhaGateList price Input$1.8$3 Output$9$15

Per 1M tokens, USD · You save ↓40%

Copy and call

curl
curl https://nezhagate.com/v1/chat/completions -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d '{"model": "kimi-k3", "messages": [{"role": "user", "content": "Hello"}]}'
Python
from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://nezhagate.com/v1")
r = client.chat.completions.create(model="kimi-k3", messages=[{"role": "user", "content": "Hello"}])
print(r.choices[0].message.content)

Replace YOUR_API_KEY with a key from your console.

Live Test · Playground

Try out Kimi K3 right here (available after login).

Input

Advanced
Web search
Memory · multi-turn

Conversation

Start a conversationType a message below to begin

About Kimi K3

Kimi K3 is the new flagship from Moonshot AI, served on NezhaGate through the OpenAI-compatible API. It is strong at long-document understanding, multi-document analysis and Chinese writing, with excellent agent and tool use. It always thinks before it answers: the reasoning comes back in message.reasoning_content and the answer in content. The sampling temperature is fixed by the model, so the gateway ignores temperature / top_p in the request instead of returning an error. Tool calling, streaming and prompt caching are supported. Set model to kimi-k3; pay-as-you-go, failed calls never billed.

Use cases

Long documents and multi-document analysis

Summaries, comparisons and Q&A over long contracts, papers or financial reports; we verified it pinpoints facts inside documents of more than 100K tokens.

Agents and tool calling

OpenAI-format tools / tool_calls for retrieval augmentation, function calls and multi-step task orchestration.

Chinese writing

Long-form articles, reports, marketing copy and polishing in natural, fluent Chinese.

Complex reasoning

It thinks before it answers, which suits maths, logic and multi-step problems; the reasoning is visible in reasoning_content.

How to choose

Pick Kimi K3 first for long documents, multi-document work and agent tasks; for everyday chat at a lower cost, DeepSeek V4.1 Flash or GLM-5.3 also fit. Switching is just the model field.

FAQ

Does the temperature parameter do anything?
Kimi K3 uses a sampling temperature fixed by the model, so the gateway ignores temperature and top_p in the request (no error is returned). The output style is decided by the model itself.
Can I turn thinking off?
No. Kimi K3 always thinks before it answers, and the reasoning tokens are billed at the output rate inside usage.completion_tokens. For shorter replies, ask for concise answers in the prompt.
What should I know about long answers?
For long content that takes more than about 2 minutes to generate, use stream=true; a non-streaming request can be cut off by a timeout. If max_tokens is below 256, the gateway raises it to 256 so the reasoning does not use up the whole budget.
Is prompt caching supported?
Yes. A repeated long prefix, such as the same long document, hits the cache, and the cached part settles at the cache rate, far below the input rate.
Am I charged for failed calls?
No. Only calls that return normally are settled, at actual usage; upstream errors and timeouts never touch your balance.

Related Models

Explore other models you can integrate.

View all →
GLM-5.3 Chat

GLM-5.3

glm-5.3

The new Z.ai flagship - coding and agents

$0.34/1M in · $1.25/1M out ↓73% View →
Qwen3.7 Max Chat

Qwen3.7 Max

qwen3.7-max

Alibaba's Qwen flagship - reasoning and coding

$1.65/1M in · $4.85/1M out ↓35% View →
Grok 4.7 Chat

Grok 4.7

grok-4.7

xAI's flagship reasoning model - long context

$0.30/1M in · $0.90/1M out ↓85% View →
DeepSeek V4.1 Flash Chat

DeepSeek V4.1 Flash

deepseek-v4.1-flash

DeepSeek's current Flash model - thinking on or off

$0.09/1M in · $0.36/1M out ↓40% View →
Claude Sonnet 4.6 Chat

Claude Sonnet 4.6

claude-sonnet-4-6

Balanced, efficient chat and coding model

$1.50/1M in · $7.50/1M out ↓50% View →
Gemini 3.1 Pro Chat

Gemini 3.1 Pro

gemini-3.1-pro

Next-generation flagship reasoning model from Google

$0.50/1M in · $3.00/1M out ↓75% View →