Moonshot Kimi

Mainland, international, and gateway access

The mainland direct base is https://api.moonshot.cn/v1 and uses a Moonshot API key. Mainland platform.kimi.com and international platform.kimi.ai isolate accounts, balances, and keys, so using the wrong site commonly returns 401. This gateway uses its own base URL and Bearer token and declares only /v1/chat/completions; provider balance, file, and token-estimation APIs are not gateway routes.

Model parameters are not interchangeable

Thinking controls differ by family: K3 uses reasoning_effort, K2.6 uses thinking, and K2.7 Code always thinks and cannot disable it. Tool-choice capabilities also differ; for example, tool_choice=required is not universal. Through this gateway, first select an enabled model and send only fields supported by that exact model.

Streaming and tool loops

Streamed tool-call arguments arrive incrementally by index and must be concatenated before JSON parsing. The application validates and executes functions and returns a role=tool message with the same tool_call_id. Preserve the full assistant message, including reasoning_content and tool_calls, across turns; new integrations should not use the deprecated function_call shape.

Errors, limits, and long requests

For 401, check the mainland/international site and key. For 429, inspect error.type to distinguish engine overload, balance, concurrency, RPM, TPM, or TPD. Retain request_id and queue, throttle, or stop retries according to the cause. Prefer streaming and explicit timeouts for long requests rather than discovering failure only at a server timeout.

Cost and quality safety

Tool definitions, context, and reasoning all consume tokens, and a large context does not guarantee retrieval quality. Cap output and concurrency, alert on usage, trim history, and add retrieval and citations for critical claims. Models, pricing, and tier limits change, so use live documentation rather than hard-coded values.