MiniMax parameter and protocol differences

For MiniMax 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
MiniMax-M3.1-Flash	Chat, Responses and Messages; thinking is always enabled, with adjustable effort
MiniMax-M3	Chat, Responses and Messages; thinking can be enabled or disabled, with different defaults by protocol

Chat Completions differences

Field	Difference
`thinking.type`	Both models default to `adaptive`; M3 can disable thinking with `disabled`, while M3.1-Flash returns `400` for `disabled`
`reasoning_effort`	M3.1-Flash supports `low/medium/high/xhigh/max`, defaulting to `max`; `none` returns `400`. M3 ignores this field
`reasoning_split`	MiniMax extension: `true` places thinking separately in `reasoning_content`; `false` puts it within `<think>` tags in `content`. It does not enable or disable thinking; M3.1-Flash does not support `false`
`temperature`, `top_p`	Values are ignored; actual values are `1` and `0.95`, respectively. If supplied, `top_p` must still satisfy `(0,1]`
`presence_penalty`, `frequency_penalty`, `logit_bias`	Values are ignored
`max_completion_tokens`	Minimum: `1`; maximum: `524288`; recommended: `131072`. Includes thinking and the final answer; the recommended value is not the default
`service_tier`	M3 supports `standard/priority`, defaulting to `standard`; M3.1-Flash does not support this field
Historical assistant messages	Replay complete tool conversations, including `reasoning_content`, `content` and `tool_calls`; simple tool calls may have no visible thinking field

Responses differences

Field	Difference
`reasoning.effort`	M3.1-Flash always reasons, supporting `low/medium/high/xhigh/max` with default `max`; `none` returns `400`. M3 defaults to no thinking; a non-`none` effort enables it without adjusting depth
`temperature`, `top_p`	Values are ignored; actual values are `1` and `0.95`, respectively. If supplied, `top_p` must still satisfy `(0,1]`
`max_output_tokens`	Thinking tokens count toward the output budget; a small budget may produce `incomplete` with no final-answer message item
`tool_choice`	Only `auto` is supported
`text.format.type`	Currently only `text` is defined
`store`	The request value is ignored; the response is fixed at `false`
`previous_response_id`, `conversation`	The request value is ignored; the response is fixed at `null`. The client must replay all historical input and output, including reasoning and tool-call items
`parallel_tool_calls`, `truncation`	Request values are ignored; response values are fixed at `true` and `disabled`, respectively
`service_tier`	M3 supports `standard/priority`, defaulting to `standard`; M3.1-Flash does not support this field
Server-side search	Beta tool; use `type: "web_search"` in `tools`

Messages differences

Field	Difference
`thinking.type`	M3.1-Flash is always enabled and only accepts `adaptive`; `disabled` returns `400`. M3 defaults to off; enable with `adaptive` or disable with `disabled`
`output_config.effort`	M3.1-Flash supports `low/medium/high/xhigh/max`, defaulting to `max`; `none` returns `400`. M3 ignores this field
`temperature`, `top_p`	Values are ignored; actual values are `1` and `0.95`, respectively. If supplied, `top_p` must still satisfy `(0,1]`
`top_k`, `stop_sequences`, `mcp_servers`, `context_management`, `container`	Values are ignored
`max_tokens`	Minimum: `1`; maximum: `524288`; recommended: `131072`; thinking tokens also count toward this limit
`tool_choice.type`	Only `auto` is supported
`service_tier`	M3 supports `standard/priority`, defaulting to `standard`; M3.1-Flash does not support this field
Historical assistant content blocks	Preserve all thinking, text and tool_use blocks and the thinking signature `signature`; simple tool calls may have no thinking block
Server-side search	Beta tool; use `type: "web_search_20250305"` and `name: "web_search"`

Cache differences

Field	Difference
Automatic prefix caching	Applies to calls with at least 512 input tokens; no cache markers are needed. Prefixes match tools, system prompts and conversation history in that order
`cache_control`	Neither model supports Anthropic's explicit cache markers
Responses `prompt_cache_key`	Can serve as a prompt-cache routing identifier

Media input differences

Input	Protocol content block	Difference
Image	`image_url`	Chat: `image_url.url` uses a public URL or Base64 Data URL
Video	`video_url`	Chat: `video_url.url` uses a public URL, Base64 Data URL or vendor `mm_file://{file_id}` reference
Image	`input_image`	Responses: `image_url` can be a URL / Base64 string or an object containing `url` and `detail`
Video	`input_video`	Responses: `video_url` can be a URL / Base64 string or an object containing `url`, `fps`, `detail` and `max_long_side_pixel`
Image	`image`	Messages: provide `url` for `source.type=url`; provide `media_type` and `data` for `source.type=base64`
Video	`video`	Messages extension: `source` supports URLs, Base64 or native video file references, with frame-sampling options

Field	Difference
`detail`	Supports `low/default/high`; default: `default`. Chat uses the relevant media object, Responses the URL object, and Messages `source`
`fps`	Video frame-sampling rate; range: `0.2～5`; default: `1`. Chat / Responses use `video_url.fps` in the video block; Messages uses `source.fps`
`max_long_side_pixel`	Limits the longest edge of an image or video frame; set it in Chat's media object, Responses' video URL object, or Messages' `source`
Media size	Chat / Messages: images up to 10 MB, URL / Base64 videos up to 50 MB, and request bodies up to 64 MB. Native Files API video uploads allow up to 512 MB

mm_file:// references come from MiniMax's native file upload API. This page does not provide an AGTCloud file upload endpoint.

Request examples

Replace MODEL_ID with the full platform model ID and use your own API Key. These examples use M3.1-Flash reasoning effort settings; see the table above for M3 behavior.

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"}],
    "thinking": {"type": "adaptive"},
    "reasoning_effort": "low",
    "max_completion_tokens": 4096
  }'