Skip to content
Docs Pydantic Docs

Moonshot AI / Kimi

Use Moonshot AI’s Kimi models through MoonshotAIProvider and the OpenAI-compatible Chat Completions API. The provider prefix is moonshotai:.

The Kimi K3.1 Pro API guide covers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Use thinking='low', thinking='high', or thinking='max' to select a documented reasoning effort, or omit thinking to use the provider’s default.

Install

Install Pydantic AI with the OpenAI SDK used by this integration:

Terminal
pip install "pydantic-ai-slim[openai]"

Configuration

Create an API key in the Moonshot Console.

You can set the MOONSHOTAI_API_KEY environment variable and use MoonshotAIProvider by name:

from pydantic_ai import Agent

agent = Agent('moonshotai:kimi-k3.1-pro')
...

Or initialise the model and provider directly:

from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIChatModel
from pydantic_ai.providers.moonshotai import MoonshotAIProvider

model = OpenAIChatModel(
    'kimi-k3.1-pro',
    provider=MoonshotAIProvider(api_key='your-moonshot-api-key'),
)
agent = Agent(model)
...

Custom base URL

MoonshotAIProvider uses https://api.moonshot.ai/v1 by default. To use the Chinese endpoint, set MOONSHOTAI_BASE_URL and an API key from the Chinese console:

Terminal
export MOONSHOTAI_API_KEY="your-moonshot-api-key"
export MOONSHOTAI_BASE_URL="https://api.moonshot.cn/v1"

You can then use the model-string configuration shown above.

Alternatively, pass base_url='https://api.moonshot.cn/v1' to MoonshotAIProvider. An explicit base_url overrides MOONSHOTAI_BASE_URL.

Thinking

Kimi K2.5, K2.6, K2.7 Code and K3 reason by default. The unified thinking setting maps onto Moonshot’s request fields:

  • thinking=False turns thinking off on Kimi K2.5 and K2.6 by sending thinking={'type': 'disabled'}. Kimi K3 can’t turn thinking off, so it reasons at its lowest effort, reasoning_effort='low'. Kimi K2.7 Code can’t turn thinking off either and has no effort setting, so thinking=False sends nothing.
  • thinking=True and the effort levels are sent as reasoning_effort (True as 'medium'). 'max' is sent as 'max' on Kimi K3, K2.6 and K2.7, and as 'xhigh' on the other Kimi models.