Doubao parameter and protocol differences

For Doubao 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
doubao-seed-2-1-lite-260915, doubao-seed-evolving, doubao-seed-2-1-pro-260915	All support Chat, Responses and Messages; thinking and tool rules are combined. Audio input is listed only within Lite's supported scope
doubao-seed-evolving	A continuously updated model ID, not the same version as a dated Pro model; keep calling configuration stable within a Session

Chat Completions differences

Field	Difference
`thinking.type`	Supports `enabled/disabled`; thinking is enabled by default
`reasoning_effort`	Default: `high`; effective levels: `low/medium/high`; `minimal` disables thinking. Compatibility values map `none` to `minimal` and `xhigh/max` to `high`
`max_tokens`	Default: `4096`; limits the final answer only, excluding thinking; not a combined budget
`max_completion_tokens`	Includes thinking and the final answer; the official Chat field range is `1～65536`. Mutually exclusive with `max_tokens`; do not use the model's advertised maximum output as this field's limit
`temperature`, `top_p`	Ranges: `[0,2]` and `[0,1]`, respectively; the fixed-temperature `1` rule from some old Seed 2.0 versions does not apply
`frequency_penalty`, `presence_penalty`	Unsupported by Seed 2.0 and later; a capability limitation, not official deprecation
`stop`, `logprobs`	The vendor does not support these by default on deep-thinking models; none of these three models is a listed exception
`reasoning_content`, `encrypted_content`	Returns a thinking summary and encrypted original by default. Replay `encrypted_content` unchanged for multi-turn tools; summaries cannot replace it. If both are supplied, encrypted content takes precedence
`stream_options.chunk_include_usage`	Default: `false`; `true` returns cumulative usage in each chunk. Do not sum cumulative counts across chunks
`stream_options.include_usage`	Controls the final total-usage chunk before completion. Lite returns request usage by default; other models require explicit enabling. Cumulative per-chunk usage and final total usage are separate settings
`response_format`	Supports `json_object/json_schema`, with the vendor's Beta designation; `strict: true` constrains the supported Schema subset, not every JSON Schema keyword

Responses differences

Field	Difference
`reasoning.effort`	Uses `minimal/low/medium/high` and compatibility mappings, with the same default as Chat; not a top-level request field
`max_output_tokens`	Includes thinking and the final answer; do not reuse Chat's answer-only budget or limits
`text.format`	Put JSON Schema configuration inside `text.format` with `name/schema/strict`, not Qwen's nested `text.format.json_schema` structure
`store`, `previous_response_id`	`store` defaults to `true`; a real response ID can continue history and original thinking. Previous `instructions` are not inherited automatically
Manual history replay	Without a response ID, replay all returned reasoning, messages and tool-call items; do not replace encrypted content with a thinking summary
`max_tool_calls`	Range: `1～10`; limits built-in tool-call rounds, not calls within each round. Best-effort, not a hard cap on all function calls
`service_tier`, `store`	In the vendor's `flex` mode, `store` must be `false`; server-side response storage cannot be used simultaneously

Messages differences

Field	Difference
`max_tokens`	Optional in the vendor API, unlike Claude's required field; the examples still explicitly set the output budget
`output_config.effort`	Controls thinking effort using model defaults and compatibility mappings; do not move Chat's top-level field into this request
`output_format`	Top-level extension supporting `text/json_object/json_schema`; put the Schema in `output_format.schema`, not Claude's `output_config.format`
`temperature`, `top_p`	Ranges: `[0,2]` and `[0,1]`, respectively, including valid `top_p: 0`
`stop_sequences`	The general field accepts a string or array, but the vendor does not support it by default for deep-thinking models. A listed field alone does not guarantee stop behavior
Historical thinking blocks	Replay returned `thinking` and `signature` completely, keeping values and block order unchanged; do not fabricate signatures or reconstruct history from summaries

Cache differences

Field	Difference
Implicit caching	Chat / Responses enable it automatically and cannot disable it; matching prefixes need at least `1024 Token` on these models, without guaranteeing hits
`prompt_cache_key`	Chat / Responses cache-routing hint; keep it stable within a conversation. The same key does not guarantee a fixed instance or a cache hit
Responses `caching`	Native explicit-cache extension enabled by `type: "enabled"`; `prefix: true` writes only a prefix without generating a response, while `false` generates and writes cache normally
`caching` combination restrictions	Cannot combine with `instructions` or non-Function `tools`; do not reuse another vendor's TTL or Session Header
Cache usage	Chat: inspect `usage.prompt_tokens_details.cached_tokens`; Responses: `usage.input_tokens_details.cached_tokens`; Messages: returned `usage.cache_read_input_tokens` and `usage.cache_creation_input_tokens`

Media input differences

Input	Protocol content block	Difference
Image	`image_url`	Chat: pass a public URL or Base64 Data URL in `image_url.url`; put detail in `image_url.detail`
Video	`video_url`	Chat: pass the video in `video_url.url`; `fps` belongs inside `video_url`, alongside `url`
PDF	`file`	Chat: choose one of `file.file_url/file_data/file_id`; file content currently supports PDF only
Audio	`input_audio`	Lite only, Chat: choose one of `input_audio.url/data/file_id`; `data` uses raw Base64 and requires `format`
Image	`input_image`	Responses: use a `image_url` string or native file ID, not Chat's nested object
Video	`input_video`	Responses: put the video address in a `video_url` string, with `fps` on the same content block
PDF	`input_file`	Responses: use `file_url/file_data/file_id`; maximum `50 MB` for Base64 or URL files
Audio	`input_audio`	Lite only, Responses: use `audio_url` for a URL / Base64 Data URL or reference a native file ID, unlike Chat's `input_audio.data`
Image	`image`	Messages: supply a public URL through `source`, or `media_type` with raw Base64 `data`
PDF	`document`	Messages: send PDF through `source` in a document block; do not reuse Chat's `file` block

Field	Difference
`fps`	Range: `0.2～5`; default: `1`; Chat and Responses field locations differ, as shown above
Lite audio	Audio files: at most `25 MB`; total audio-only input per request: at most `120 minutes`. Use Chat / Responses for audio input
Evolving media	Supports images, video visuals and documents; not audio or video soundtracks
Native file ID	Comes from Ark Files API and is associated with the upload project and file type; this page does not promise an equivalent AGTCloud upload endpoint

Request examples

Replace MODEL_ID with the full platform model ID 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
  }'