Qwen parameter and protocol differences

For Qwen 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
qwen3.8-max-0902, qwen3.8-flash, qwen3.8-max	All support Chat, Responses and Messages; shared rules are combined, with media endpoint scope specified below

Chat Completions differences

Field	Difference
`enable_thinking`, `reasoning_effort`	Thinking is enabled by default; effective levels: `low/medium/xhigh`; default: `xhigh`. `minimal` maps to `low`, `high/max` to `xhigh`, and `none` disables thinking
`thinking_budget`	Cannot be set together with `reasoning_effort`; doing so returns an error. When effort alone is set, `low/medium/xhigh` correspond to thinking budgets `4096/16384/262144`, respectively
`max_tokens`, `max_completion_tokens`	`max_tokens` limits only the final answer and is marked for upcoming deprecation. For new integrations, use `max_completion_tokens` to budget thinking and the answer together
`temperature`	Range: `[0,2)`; thinking-mode defaults: `1.0` for text, `0.6` for visual input; values below `0.6` become `0.6`. Non-thinking default: `0.7`
`top_p`, `top_k`	`top_p` range: `(0,1]`; thinking / non-thinking defaults: `0.95/0.8`. Extension `top_k` is a nonnegative integer, default `20`; `null` or values above `100` disable Top-K
`repetition_penalty`	Qwen extension; must exceed `0`; default: `1.0`; distinct from OpenAI's two penalty fields
`preserve_thinking`	Enabled by default; replay assistant thinking unchanged in its original `reasoning_content` field, not in `content`. Missing thinking does not itself cause an error
`tool_choice`	Supports `auto/none`; thinking mode disallows `required` or a named function. Non-thinking mode allows a named function, but `required` still does not guarantee a tool call
`tool_stream`	Default: `false`; controls incremental complex tool arguments only with `stream=true`. Max models support text scenarios only; Flash supports all modalities
`response_format`	Supports `json_object/json_schema`; the former requires a JSON keyword in the prompt; the latter defines structure through `response_format.json_schema`. Thinking need not be disabled
`enable_search`, `search_options`	`enable_search` defaults to `false`; `search_options.forced_search` can force search. These models use strategies `turbo/max`. With `include_tool_usage=true`, actual search-call statistics appear in `usage.x_tools`

Responses differences

Field	Difference
`reasoning.effort`	Supports `none/low/medium/xhigh`; default: `xhigh`; `minimal` maps to `low`, and `high/max` to `xhigh`
`enable_thinking`	Compatibility switch scheduled for removal; the vendor recommends `reasoning.effort`. If both are supplied, `reasoning.effort` takes precedence
`max_output_tokens`	Minimum: `16`; includes thinking and the final answer
`text.format`	Supports `text/json_object/json_schema`; JSON Schema settings go inside `text.format.json_schema`, containing `name/schema/strict`, unlike the original protocol's flat structure
`tool_choice`	Supports `auto/none/required` and an `allowed_tools` object; `required` is available only with exactly one tool in the list
`tools`	Alongside functions, supports built-ins such as `web_search`, `web_extractor` and `code_interpreter`. Web extraction requires search; the code interpreter requires thinking
`store`, `previous_response_id`, `conversation`	Supports server-side history; `store` defaults to `true`, and response IDs last `7 days`. The latter two fields are mutually exclusive
`background` and other extension parameters	Background asynchronous execution is unsupported; OpenAI parameters not listed in the vendor's Responses Reference are ignored

Messages differences

Field	Difference
`thinking.type`	Supports `enabled/disabled`, with thinking enabled by default; do not apply Claude's `adaptive` configuration
`output_config.effort`	Uses `low/medium/xhigh`, defaulting to `xhigh`; `high/max` maps to `xhigh`. Do not reuse all Chat compatibility mappings
`thinking.budget_tokens`	Marked for upcoming deprecation; use `output_config.effort` for new integrations. With the old budget configuration, `max_tokens` must exceed that budget
`max_tokens`	Required; on Qwen3.8, limits thinking and the final answer together
`temperature`, `messages[].role`	Temperature range: `[0,2)`; roles also allow `system` in addition to `user/assistant`
`tool_result.content`	Return tool results as strings; serialize objects or arrays first rather than using Claude's content-block arrays
`output_config.format`	Supports strict JSON Schema; use `type: "json_schema"`, put the Schema in `schema`, and set `additionalProperties` to `false`
`web_search` in `tools`	Search tool name: `web_search`; use a dated type such as `web_search_20250305`. HTTP / SDK calls must also add client identifier `x-anthropic-billing-header: cc_entrypoint=cli;` to `system`

Cache differences

Field	Difference
Implicit caching	Enabled automatically; matching prefixes need at least `1024 Token` to qualify, without guaranteeing a hit
Chat / Messages explicit caching	Mark content blocks with `cache_control: { type: "ephemeral" }`; minimum: `1024 Token`; valid for `5 minutes`, refreshed on a hit; at most `4` markers
Explicit cache locations	System prompts must use content-block arrays. Tool definitions contribute to the prefix, but markers placed on tool definitions alone are ignored. Do not reuse other TTLs or cache-breakpoint fields from the general Reference
Responses Session cache	Enable through Header `x-dashscope-session-cache: enable`; default: `disable`. Disabling Session caching does not disable automatic implicit caching
Cache usage	Chat: inspect `usage.prompt_tokens_details.cached_tokens`; Responses: `usage.input_tokens_details.cached_tokens`; Messages: `usage.cache_read_input_tokens` and `usage.cache_creation_input_tokens`

Media input differences

Input	Protocol content block	Difference
Image	`image_url`	Chat: `image_url.url` supports public URLs or Base64 Data URLs
Video	`video_url`	Chat extension: `video_url.url` supports public URLs or Base64 Data URLs; `fps` belongs on the content block alongside `video_url`
Frame list	`video`	Chat extension: pass chronological images in a `video` array; at least `4` frames
Image	`input_image`	Responses: `image_url` is a public URL or complete Base64 Data URL string, not Chat's nested object
PDF	`input_file`	The vendor explicitly lists qwen3.8-max and qwen3.8-flash for Responses. Use `file_url` for a public PDF up to `100 MB`; do not apply OCR-specific parameters to these models
Image	`image`	Messages: provide `url` for `source.type=url`; provide `media_type` and raw Base64 `data` for `base64`
Video	`video`	Messages extension: `source` contains a public URL, or `media_type` and raw Base64 `data`; this is not a native Claude content block

Field	Difference
Chat image controls	Content-block `min_pixels/max_pixels` adjust image pixels; with top-level `vl_high_resolution_images=true`, `max_pixels` no longer applies
Image size	Public images: at most `20 MB`. Images before Base64 encoding: at most `20 MB`. Encoded Data URLs: at most `20 MB` in Chat or `32 MB` in Messages
Video limits	Public videos: at most `2 GB`; Base64 strings: less than `10 MB`. Only visuals are processed, not the video's audio
Protocol scope	These models do not use Omni audio endpoints; Responses `input_audio/input_video` are currently Omni-only. Do not move Chat / Messages video blocks into Responses

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
  }'