Kimi parameter and protocol differences

For Kimi models, adjust your chosen protocol's parameters using this page. See Chat Completions, Responses and Messages for full request and response structures.

Applicable models

Model	Applies to
kimi-k3.1-pro, kimi-k3.1-pro-highspeed, kimi-k3	Shared parameter rules for the three protocols below; the applicable models for specific extensions are identified in each row

Chat Completions differences

Field	Difference
`temperature`, `top_p`	Fixed at `1.0` and `0.95`, respectively; preferably omit them. Other values return an error
`n`, `presence_penalty`, `frequency_penalty`	Fixed at `1`, `0` and `0`, respectively; other values return an error
`reasoning_effort`	Supports `low/high/max`; default: `max`. Thinking and retained thinking are always enabled; do not use the K2.x `thinking` switch
`max_completion_tokens`	Default: `131072`; maximum: `1048576`. The output budget is reduced automatically if input plus output budget exceeds the context window; the input itself must still fit
`stop`	At most 5 strings, each no longer than 32 bytes
`tool_choice`	Supports `auto/none/required`; the object form specifying a function is incompatible with enabled thinking and returns `400`
`messages[].reasoning_content`	Replay complete assistant messages unchanged in multi-turn conversations and tool calls, including thinking and `tool_calls`
`messages[].partial`	Kimi Partial Mode extension: set `true` on the final assistant message to continue from the supplied prefix; do not combine it with `response_format.type=json_object`
`messages[].tools`	Dynamic tool extension, documented for kimi-k3.1-pro and kimi-k3: append a system message containing the full tool definitions and no `content`. Tools become visible from that position; retain this message in later requests

Responses differences

Field	Difference
`reasoning.effort`	Supports `low/high/max`; default: `max`
`max_output_tokens`	Default: `131072`; maximum: `1048576`
`tool_choice`	Only `auto` is supported
`tools[].type`	Only `function/namespace/custom/web_search` is supported; the `custom` tool name must be `apply_patch`
`text.format`	Use `type: "json_schema"` and `schema`; `name` defaults to `output` when omitted
Conversation state fields in the response	`store` and `background` are fixed at `false`; `previous_response_id` and `conversation` at `null`. Replay historical input and output on the client rather than relying on a server-side response chain

Messages differences

Field	Difference
`max_tokens`	Required; minimum: `1`; `0` is not allowed
`output_config.effort`	Supports `low/high/max`; default: `max`
`tool_choice.type`	Supports `auto/any/none`; default: `auto`
`stop_sequences`	At most 5 strings, each no longer than 32 bytes
Historical `thinking` blocks	Replay them unchanged with `signature` in the next assistant message
`output_config.format`	Structured output requires `type: "json_schema"` and `schema`

Cache differences

Field	Difference
Chat / Responses `prompt_cache_options`	`mode` only supports `implicit`; `ttl` supports `5m/1h`, defaulting to `5m`. Omitting the configuration still writes a 5-minute cache by default
Chat / Responses `prompt_cache_breakpoint`	Explicit breakpoints are unsupported; including one in content returns `400`
Messages `cache_control`	Only the request-level configuration takes effect; markers within `messages` are ignored. `type` is fixed at `ephemeral`; `ttl` supports `5m/1h`, defaulting to `5m`. Without it, the cache is still read and written with a 5-minute TTL
Reasoning effort	Changing the effort level breaks prefix cache hits; select it at the start of the conversation

Media input differences

Input	Protocol content block	Difference
Image	`image_url`	Chat: `image_url.url` uses a Base64 Data URL or a vendor `ms://<file_id>` reference; ordinary public image URLs are unsupported
Video	`video_url`	Chat: use a Base64 Data URL or an `ms://<file_id>` reference to an uploaded video
Image	`input_image`	Responses: `image_url` uses a Base64 Data URL; public HTTP(S) URLs are unsupported
Image	`image`	Messages: provide `media_type` and `data` for `source.type=base64`; for `source.type=url`, supply the vendor `ms://<file_id>` reference in `source.url`
SVG	Image input	SVG images are unsupported, including Base64. To analyze SVG, send its XML source as text instead

ms:// is Kimi's native file reference format and requires uploading the file to Kimi first. This page does not provide an AGTCloud file upload endpoint.

Request examples

Replace MODEL_ID with the full model ID available at your selected protocol endpoint, and use your own API Key.

curl 'https://api.agtcloud.ai/v1/chat/completions' \
  -H 'Authorization: Bearer your-api-key' \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "MODEL_ID",
    "messages": [{"role": "user", "content": "Explain how to troubleshoot an API timeout in three points"}],
    "reasoning_effort": "low",
    "max_completion_tokens": 4096,
    "prompt_cache_options": {"mode": "implicit", "ttl": "1h"}
  }'