Qwen3 VL 32B Instruct

alibaba/qwen3-vl-32b-instruct — a llm model by Alibaba. Call it through the gateway’s standard OpenAI-compatible door using the catalog slug; the full parameter schema is reproduced in full below.

Catalog slug	alibaba/qwen3-vl-32b-instruct
Kind	llm
Vendor	Alibaba
Context window	1,000,000 tokens
Native model id	qwen3-vl-32b-instruct

How to call

Through the Sociaro SDK / OpenAI-compatible door

from sociaro_ai import Sociaro

client = Sociaro(api_key="gw_live_...")   # base_url defaults to https://api.sociaro.com

resp = client.chat.completions.create(
    model="alibaba/qwen3-vl-32b-instruct",          # always the catalog slug
    messages=[{"role": "user", "content": "Hello"}],
    # any parameter from the schema below is passed through unchanged
)
print(resp.choices[0].message.content)

TypeScript: client.chat.completions.create({ model: "alibaba/qwen3-vl-32b-instruct", messages: [...] }) — see the TypeScript SDK.

Through Alibaba DashScope’s native API (advanced)

Optional: send Alibaba DashScope’s native request to the /alibaba prefix, using the native model id qwen3-vl-32b-instruct. The gateway injects the upstream Alibaba DashScope credential and forwards the body unchanged (catalog, pricing and entitlements still apply).

curl https://api.sociaro.com/alibaba/<native-path> \
  -H "Authorization: Bearer $SOCIARO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "qwen3-vl-32b-instruct", ... }'

The exact path, request body and response shape are documented in full in the schema below.

Parameters — full schema

The complete, one-to-one API schema for this model, transcribed from the official source. Nothing omitted. Examples in the schema may use a representative model of the family — for this model use the slug alibaba/qwen3-vl-32b-instruct (or the native id qwen3-vl-32b-instruct for the native API) as shown in How to call above.

0. Family coverage map

ONE door, ONE parameter contract. All slugs are invoked through either:

• OpenAI-compatible: POST /compatible-mode/v1/chat/completions, OR
• Native DashScope: POST /api/v1/services/aigc/multimodal-generation/generation

The parameter set is identical across the family; only defaults / numeric ranges (pixel limits, thinking defaults) differ per sub-family.

Model group (slugs)	Vision	OCR tasks	enable_thinking	top_k def	presence_penalty def	px/token	hi-res cap
qwen3-vl-plus, qwen3-vl-flash (hybrid)	yes	no	toggleable (def per Supported-models table)	20	1.5 (non-thinking)	32×32	16,777,216
qwen3-vl-235b-a22b-instruct / -thinking (close-source)	yes	no	instruct=toggleable; -thinking = always-on (cannot disable)	20	1.5 (non-thinking)	32×32	16,777,216
qwen3-vl-* other open-source (8b/30b-a3b/32b × instruct/thinking)	yes	no	-thinking always-on; -instruct toggleable	20	1.5 (non-thinking)	32×32	16,777,216
qwen3.7-max-vision	yes	no	toggleable (def per Supported-models table)	20	1.5 (non-thinking)	32×32	16,777,216
qwen-vl-ocr, qwen-vl-ocr-latest, qwen-vl-ocr-2025-11-20, qwen-vl-ocr-2025-08-28, qwen-vl-ocr-2025-04-13, qwen-vl-ocr-2024-10-28	yes (OCR-specialized)	yes (ocr_options, enable_rotate)	n/a	—	1.5 rec. for OCR extraction	32×32 for ocr / latest / 2025-11-20; 28×28 for older snapshots	—

NOTE on adjacent families that ride the SAME door (NOT part of qwen-vl, but params appear because the contract is shared): Qwen3.7/3.6/3.5 text series, Qwen3.5-Omni / Qwen-Omni (modalities/audio), DeepSeek-V4 (reasoning_effort), MiniMax-M3 (thinking object), GLM, Kimi. Pixel rows for these are kept in the value tables for completeness because the official pixel tables enumerate them together.

Critical per-model differences (read first)

1. OCR built-ins (ocr_options, enable_rotate) exist ONLY on qwen-vl-ocr* and ONLY via native DashScope SDK/HTTP. In OpenAI-compatible mode you must hand-write the task prompt into the text part.
2. Pixel-per-token = 32×32 for the Qwen3-VL series, qwen3.7-max-vision, and the Qwen3-VL-architecture OCR snapshots (qwen-vl-ocr / -latest / -2025-11-20). = 28×28 for the older OCR snapshots (-2025-08-28 / -2025-04-13 / -2024-10-28). This drives the hi-res cap (16,777,216 vs 12,845,056) and token counts.
3. Thinking: -thinking slugs always reason and ignore enable_thinking=false. Hybrid qwen3-vl-plus/-flash and qwen3.7-max-vision honor enable_thinking. Per-model enable_thinking defaults are listed in the “Supported models” table referenced from the chat-completions page.
4. OCR max_tokens: defaults to 4096 for qwen-vl-ocr, qwen-vl-ocr-2025-04-13, qwen-vl-ocr-2025-08-28; raisable to 4097–8192 only by contacting the commercial manager.
5. Audio-from-video is NOT supported by any Qwen-VL model (only Qwen-Omni). modalities/audio output params apply only to Qwen-Omni.

---

1. Request body — shape (OpenAI-compatible mode)

{
  "model": "qwen3-vl-plus",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "image_url",
          "image_url": { "url": "https://.../img.jpg" },
          "min_pixels": 65536,
          "max_pixels": 2621440 },
        { "type": "text", "text": "Describe this image." }
      ]
    }
  ],
  "stream": false,
  "temperature": 0.7,
  "max_tokens": 2048,
  "vl_high_resolution_images": false,
  "enable_thinking": false,
  "thinking_budget": 81920
}

Non-standard OpenAI fields (vl_high_resolution_images, enable_thinking, thinking_budget, top_k, repetition_penalty, enable_search, search_options, tool_stream, preserve_thinking, reasoning_effort, enable_code_interpreter, thinking): in Python SDK pass via extra_body={...}; in Node.js SDK as top-level fields; in curl/HTTP (compatible-mode) as top-level body fields; in native DashScope HTTP inside "parameters".

1b. Request body — shape (native DashScope mode)

{
  "model": "qwen3-vl-plus",
  "input": {
    "messages": [
      { "role": "user",
        "content": [
          { "image": "https://.../img.jpg", "min_pixels": 65536, "max_pixels": 2621440 },
          { "text": "Describe this image." }
        ] }
    ]
  },
  "parameters": {
    "vl_high_resolution_images": false,
    "incremental_output": false,
    "result_format": "message",
    "ocr_options": { "task": "advanced_recognition" }
  }
}

Native mode: content keys are image / video / text (no image_url/video_url/type); all generation params live in the parameters object.

---

2. Top-level request parameters (complete table)

name	type	required	default	allowed / range	description
model	string	yes	—	model slug	Model name (qwen3-vl-*, qwen-vl-ocr*, qwen3.7-max-vision).
messages	array	yes	—	System / User / Assistant / Tool message objects	Conversation history in chronological order. content structure see §3.
stream	boolean	no	false	true / false	Stream output in chunks. Recommended true for long outputs (else 300s timeout risk).
stream_options	object	no	—	{ "include_usage": bool }	Effective only when stream=true. include_usage (bool, def false) adds a final usage chunk.
temperature	float	no	model-dependent (console)	[0, 2)	Sampling temperature. Set ONLY temperature OR top_p.
top_p	float	no	model-dependent (console)	(0, 1.0]	Nucleus sampling. Set ONLY top_p OR temperature.
top_k	integer	no	Qwen3-VL series, qwen3.7-max-vision & “all other models”: 20; Qwen3-Omni-Flash: 50	integer ≥ 0; null or >100 → top_k disabled (top_p only)	Number of candidate tokens. Non-OpenAI → extra_body. DeepSeek/Kimi/MiniMax do NOT support top_k.
repetition_penalty	float	no	model-dependent	> 0, no strict upper bound; 1.0 = no penalty	Penalty on repeated consecutive sequences. Non-OpenAI → extra_body.
presence_penalty	float	no	Qwen3-VL (non-thinking) / qwen3.7-max-vision: 1.5; thinking modes of qwen3-8b…235b etc.: 0.5; all other: 0.0	[-2.0, 2.0]	Presence penalty (independent of frequency). Positive → less repetition. 1.5 recommended for OCR text extraction.
response_format	object	no	{"type":"text"}	{"type":"text"} / {"type":"json_object"}	Response format. With json_object you MUST explicitly request JSON in the prompt, else error.
max_tokens	integer	no (Deprecated)	model max output (console)	integer	Max response tokens; on hit finish_reason=length. Does NOT limit CoT. Replacement → max_completion_tokens. (OCR qwen-vl-ocr/-2025-04-13/-2025-08-28: default 4096, raisable to 8192 by contacting commercial manager.)
max_completion_tokens	integer	no	model max output (console)	integer	Max response tokens INCLUDING CoT. Recommended for reasoning models.
vl_high_resolution_images	boolean	no	false	true / false	Increases the max input-image pixel limit to the pixel value of 16384 tokens. true → fixed-resolution strategy, max_pixels is IGNORED (image downscaled to the cap). Cap = 16,777,216 (Qwen3-VL series / qwen3.7-max-vision / OCR snapshots on the Qwen3-VL architecture); = 12,845,056 (older 28×28 OCR snapshots). false → limit follows max_pixels. Non-OpenAI → extra_body. Applies to image inputs.
n	integer	no	1	1..4	Number of response candidates. Qwen3 (non-thinking) only; with tools → n=1. Increases output tokens.
enable_thinking	boolean	no	varies per model (see Supported-models table)	true / false	Enables thinking for hybrid models. With true, CoT in reasoning_content. Non-OpenAI → extra_body. No effect on -thinking (always-on) models. DeepSeek-V4 uses reasoning_effort; MiniMax-M3 uses thinking instead.
thinking_budget	integer	no	model max CoT length	integer	Token cap on reasoning; on exceed → CoT truncated and answer emitted immediately. Applies to Qwen3-VL. Non-OpenAI → extra_body.
reasoning_effort	string	no	high	high / max (low,medium→high; xhigh→max)	Reasoning intensity for DeepSeek-V4 (deepseek-v4-pro / -flash). Non-OpenAI → extra_body. (Listed for door completeness; not a Qwen-VL knob.)
thinking	object	no	{"type":"adaptive"}	type ∈ adaptive / disabled	Thinking mode for MiniMax/MiniMax-M3 (Xiyu direct) — alternative to enable_thinking. Non-OpenAI → extra_body. (Door-shared.)
preserve_thinking	boolean	no	false	true / false	Add prior assistant reasoning_content into the input (billed). Non-OpenAI → extra_body. Supported on qwen3.7-max/-2026-05-20+ and later snapshots.
seed	integer	no	—	[0, 2^31−1]	Seed for reproducibility.
logprobs	boolean	no	false	true / false	Return log-probabilities. Supported on Qwen-plus snapshots (excluding stable), incl. qwen3-vl-plus / qwen3-vl-flash. reasoning_content has no logprobs.
top_logprobs	integer	no	0	0..5	Top candidates per step. Effective only when logprobs=true.
stop	string | array	no	—	string / array of strings OR token_id; do not mix strings and token_id in one array	Stop words/tokens — generation halts.
tools	array	no	—	array of {type:"function", function:{name, description, parameters}}	Tools for function calling. name ≤ 64 chars, [A-Za-z0-9_-].
tool_choice	string | object	no	auto	auto / none / {"type":"function","function":{"name":"..."}}	Tool selection policy. Forcing a specific tool is NOT supported in thinking mode.
parallel_tool_calls	boolean	no	false	true / false	Enable parallel tool calling.
tool_stream	boolean	no	false	true / false	Stream complex tool arguments (only when stream=true; Qwen & GLM series). Non-OpenAI → extra_body.
enable_search	boolean	no	false	true / false	Enable web search (increases token consumption; may need forced_search). Non-OpenAI → extra_body.
search_options	object	no	—	see §2a	Web search strategy (effective when enable_search=true). Non-OpenAI → extra_body.
enable_code_interpreter	boolean	no	false	true / false	Enable the Code Interpreter feature. Non-OpenAI → extra_body. (Door-shared.)
ocr_options	object	no	—	{ "task": <enum>, "task_config": {...} }	Native DashScope SDK/HTTP only (absent in OpenAI mode). Built-in OCR tasks. See §4.
modalities	array	no	["text"]	["text"] / ["text","audio"]	Output modalities. Qwen-Omni only (NOT Qwen-VL).
audio	object	no	—	{voice (required), format (required, only "wav")}	Output audio voice/format. Qwen-Omni only, requires modalities=["text","audio"].

2a. search_options sub-schema

property	type	required	default	allowed	description
forced_search	boolean	no	false	true / false	Force web search (only when enable_search=true).
search_strategy	string	no	turbo	turbo / max / agent / agent_max	Search scale. turbo=balanced; max=multi-engine; agent=multi-round (qwen3.5-plus/flash, qwen3-max snapshots, qwen3.5-omni-plus/flash); agent_max=agent+web extraction (thinking mode of qwen3-max / qwen3-max-2026-01-23 only).
enable_search_extension	boolean	no	false	true / false	Domain-specific search (only when enable_search=true). Non-OpenAI → extra_body.

2b. Native-DashScope parameters (not on the OpenAI-compatible page)

These live in the parameters object on the native DashScope generation API (and in the Python/Java/Node native SDKs).

name	type	default	scope	description
incremental_output	boolean	false	native only	Native streaming: emit incremental deltas instead of the full cumulative text each chunk. The OpenAI-compatible equivalent is stream + stream_options (deltas are returned natively in compatible mode).
result_format	string	message	native only	Shape of the returned data: message (OpenAI-style messages) or text. Must be message when using tools (function calling).
vl_enable_image_hw_output	boolean	false	native only	Return the height/width of the scaled input image. Casing differs by SDK: vl_enable_image_hw_output in the HTTP parameters object and the Python SDK; vlEnableImageHwOutput in the Java SDK (≥ 2.20.8).
max_frames	integer	model-dependent	native only	Cap on the number of extracted video frames. Not supported on the OpenAI-compatible path — that path uses each model’s default. Default/max are model-dependent (8000 for the top qwen3.x series).

---

3. messages[].content structure (multimodal input)

content is a string (text only) or an array of part objects.

field	type	required	values / description
type	string	yes (OpenAI mode)	text / image_url / input_audio / video / video_url. (Native mode: no type; the part key IS the type.)
text	string	when type=text	Input text.
image_url	object	when type=image_url	{ "url": <string, required> }. url = HTTP(S) or Base64 Data URL.
input_audio	object	when type=input_audio	{ "data": <url/base64, required>, "format": <"mp3"/"wav", required> }. (Omni.)
video	array	when type=video	Image list (frames), e.g. ["https://f1.jpg", ...].
video_url	object	when type=video_url	{ "url": <required> } + optional fps. Only some Qwen-VL accept a video file.
fps	float	no (inside video/video_url)	[0.1, 10], default 2.0. Frames extracted per second (1 frame each 1/fps s); also informs the model of inter-frame interval. (MiniMax-M3: range [0.2, 5], default 1.)
min_pixels	integer	no (inside image/video part)	Lower pixel threshold; below → upscale. Defaults see §3a.
max_pixels	integer	no (inside image/video part)	Upper pixel threshold; above → downscale. Defaults see §3a.
total_pixels	integer	no (inside video/image-list part)	Caps total pixels across ALL extracted frames (px/frame × frame count); per-frame stays within [min_pixels, max_pixels]. Lower it for long videos to cut tokens. Defaults/max see §3a.
enable_rotate	boolean	no (inside image part, OCR)	false default. Auto image-rotation correction. Native DashScope SDK/HTTP only, qwen-vl-ocr.
cache_control	object	no (on a system- or user-message content block)	{ "type": "ephemeral" } — marks the block for explicit context caching. Gated to a supported-model subset. Available on the native path AND the OpenAI-/Anthropic-compatible paths. Min 1024 tokens per cached block, up to 4 markers per request, 5-minute TTL.

Native content keys: {"image": <url>}, {"video": <url|list>}, {"text": <str>} (instead of image_url/video_url/type).

3a. min_pixels / max_pixels / total_pixels / vl_high_resolution_images per-model values

min_pixels — IMAGE input (default = minimum):

Family	default = min
Qwen3.7 / Qwen3.6 / Qwen3.5 / Qwen3-VL / qwen3.7-max-vision	65536
Qwen3.5-Omni	24576

min_pixels — VIDEO file / image-list input:

Family	default	min
Qwen3.7/3.6/3.5-Omni/3.5/Qwen3-VL (close-source+open-source), qwen3.7-max-vision	65536	4096

max_pixels — IMAGE input, vl_high_resolution_images=false:

Family	default	max
Qwen3.7/3.6/3.5 / Qwen3-VL / qwen3.7-max-vision	2621440	16777216
Qwen3.5-Omni	1310720	16777216

vl_high_resolution_images=true → max_pixels invalid, fixed cap:

Family	fixed cap (px)	px/token
Qwen3.7/3.6/3.5-Omni/3.5 / Qwen3-VL / qwen3.7-max-vision / OCR snapshots on the Qwen3-VL arch	16777216 (16384×32×32)	32×32
older 28×28 OCR snapshots (qwen-vl-ocr-2025-08-28 / -2025-04-13 / -2024-10-28)	12845056 (16384×28×28)	28×28

total_pixels — VIDEO file / image-list (default = max, per token math at 32×32):

Family	default = max	tokens
Qwen3.7 / Qwen3.6 / Qwen3.5 series	819200000	800,000 image tokens
Qwen3.5-Omni	184549376	180,224 image tokens
close-source Qwen3-VL (qwen3-vl-235b-a22b-thinking / -instruct), qwen3.7-max-vision	134217728	131,072 image tokens
other open-source Qwen3-VL (8b / 30b-a3b / 32b)	67108864	65,536 image tokens

Image-token formula: tokens = h_bar*w_bar / token_pixels + 2 (h_bar, w_bar = scaled dims, multiples of 32 or 28).

---

4. OCR — ocr_options (native DashScope SDK/HTTP only; qwen-vl-ocr*)

"ocr_options": {
  "task": "key_information_extraction",
  "task_config": {
    "result_schema": {
      "Ride Date": "ride date, format YYYY-MM-DD",
      "Invoice Code": "invoice code",
      "Invoice Number": "invoice number"
    }
  }
}

task value	Purpose	Built-in prompt (auto-injected)	Output format
advanced_recognition	Hi-precision recognition + text-line localization	”Locate all text lines and return the coordinates of the rotated rectangle ([cx, cy, width, height, angle]).”	Plain text OR JSON in ocr_result.words_info[] with text, location ([x1,y1..x4,y4], 4 vertices TL→TR→BR→BL, origin top-left), rotate_rect ([center_x, center_y, width, height, angle], angle ∈ [-90, 90]). Requires versions later than qwen-vl-ocr-2025-08-28 or latest.
key_information_extraction	KIE from receipts/forms/certs	”Assume you are an information extraction expert. You are given a JSON schema. Fill the value part of this schema with information from the image…”	JSON object in ocr_result.kv_result. With task_config.result_schema (key→hint, up to 3 nested layers) → custom fields; without schema → all detected fields.
table_parsing	Table parsing	”…converting tables from a synthetic image into HTML. Transcribe each table using <tr> and <td> tags…merged cells accurately represented…”	Text in HTML (<tr>/<td>, merged cells).
document_parsing	Document/scan/PDF-as-image parsing	”…transcribe the text, tables, and equations…into LaTeX format without modification…”	Text in LaTeX.
formula_recognition	Formula recognition	”Extract and output the LaTeX representation of the formula from the image, without any additional text or descriptions.”	Text in LaTeX.
text_recognition	General text (zh/en)	“Please output only the text content from the image without any additional descriptions or formatting.”	Plain text.
multi_lan	Multilingual text (ar, fr, de, it, ja, ko, pt, ru, es, vi)	“Please output only the text content from the image without any additional descriptions or formatting.”	Plain text.

task_config properties: result_schema (object) — JSON field template (key → description/hint) for KIE custom-field extraction; supports up to 3 nested layers; omit to extract all fields.

In the OpenAI-compatible SDK these built-in tasks are NOT available as a parameter — hand-write the task’s fixed prompt into text. Default prompt when text is empty: "Please output only the text content from the image without any additional descriptions or formatting."

enable_rotate (boolean, default false) — auto image-rotation correction; passed inside the image content-item (NOT inside ocr_options); native DashScope SDK/HTTP only.

OCR snapshots (verified on the OCR page): qwen-vl-ocr, qwen-vl-ocr-latest, qwen-vl-ocr-2025-11-20 (Qwen3-VL architecture, 32×32 px/token), qwen-vl-ocr-2025-08-28, qwen-vl-ocr-2025-04-13, qwen-vl-ocr-2024-10-28. OCR max_tokens: qwen-vl-ocr / -2025-04-13 / -2025-08-28 default 4096, raisable to 4097–8192 only via commercial manager.

---

5. Response — shape (chat.completion, non-streaming, OpenAI mode)

{
  "id": "chatcmpl-6ada9ed2-...",
  "object": "chat.completion",
  "created": 1735120033,
  "model": "qwen3-vl-plus",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "These images show a person and a dog on a beach...",
        "reasoning_content": null,
        "tool_calls": null,
        "refusal": null
      },
      "finish_reason": "stop",
      "logprobs": null
    }
  ],
  "usage": {
    "prompt_tokens": 1277,
    "completion_tokens": 81,
    "total_tokens": 1358,
    "prompt_tokens_details": { "image_tokens": 2497, "text_tokens": 209, "cached_tokens": 2048 },
    "completion_tokens_details": { "text_tokens": 81 }
  },
  "system_fingerprint": null,
  "service_tier": null
}

response field	type	description
id	string	Unique request identifier.
object	string	Always chat.completion.
created	integer	Unix timestamp (s).
model	string	Model used.
choices[]	array	Output candidates (1..n).
choices[].index	integer	Index in choices.
choices[].finish_reason	string	stop (natural / stop word) / length (max_tokens hit) / tool_calls (tool invoked).
choices[].logprobs	object|null	content[] with token, bytes, logprob, top_logprobs[].
choices[].message.role	string	Always assistant.
choices[].message.content	string	Answer text.
choices[].message.reasoning_content	string|null	Chain of thought (thinking mode).
choices[].message.tool_calls[]	array	{id, type:"function", function:{name, arguments(JSON string)}, index}.
choices[].message.refusal / audio / function_call	—	null (function_call deprecated).
usage.prompt_tokens / completion_tokens / total_tokens	integer	Token accounting.
usage.prompt_tokens_details.image_tokens / text_tokens / cached_tokens	integer	Input breakdown (image tokens, cached tokens).
usage.completion_tokens_details.text_tokens	integer	Output breakdown.
system_fingerprint / service_tier	string|null	Currently null.

5b. Response — shape (native DashScope)

{
  "output": {
    "choices": [
      { "finish_reason": "stop",
        "message": { "role": "assistant",
          "content": [ { "text": "..." } ] } }
    ]
  },
  "usage": { "output_tokens": 81, "input_tokens": 1277, "image_tokens": 2497,
             "input_tokens_details": { "text_tokens": 33, "image_tokens": 1377 },
             "output_tokens_details": { "text_tokens": 38 }, "total_tokens": 1358 },
  "request_id": "ccf845a3-..."
}

OCR advanced_recognition: inside output.choices[0].message.content[0], alongside text, there is ocr_result.words_info[] (see §4). KIE returns ocr_result.kv_result.

---

6. Minimal example (OpenAI-compatible, single image)

from openai import OpenAI
import os
client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
    model="qwen3-vl-plus",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "https://.../img.jpg"}},
        {"type": "text", "text": "What is in this image?"},
    ]}],
)
print(completion.choices[0].message.content)

curl -X POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"qwen3-vl-plus","messages":[{"role":"user","content":[
        {"type":"image_url","image_url":{"url":"https://.../img.jpg"}},
        {"type":"text","text":"What is in this image?"}]}]}'

7. Full example (hi-res + thinking + video + per-image pixel limits + total_pixels, streaming)

completion = client.chat.completions.create(
    model="qwen3-vl-235b-a22b-thinking",
    messages=[{"role": "user", "content": [
        {"type": "image_url",
         "image_url": {"url": "https://.../doc.jpg"},
         "min_pixels": 65536,
         "max_pixels": 16384 * 32 * 32},
        {"type": "video_url",
         "video_url": {"url": "https://.../clip.mp4"},
         "fps": 2,
         "min_pixels": 65536,
         "max_pixels": 2621440,
         "total_pixels": 134217728},
        {"type": "text", "text": "Solve the problem in the image and summarize the video."},
    ]}],
    stream=True,
    temperature=0.7,
    top_p=0.8,           # NOTE: official guidance — set only temperature OR top_p
    seed=1234,
    max_completion_tokens=4096,
    stream_options={"include_usage": True},
    extra_body={
        "vl_high_resolution_images": True,   # max_pixels ignored
        "enable_thinking": True,             # -thinking models always reason regardless
        "thinking_budget": 81920,
        "top_k": 20,
        "repetition_penalty": 1.05,
    },
)
for chunk in completion:
    if not chunk.choices:
        print(chunk.usage)
    else:
        d = chunk.choices[0].delta
        if getattr(d, "reasoning_content", None):
            print(d.reasoning_content, end="")
        elif d.content:
            print(d.content, end="")

8. Full example (OCR built-in KIE task with custom schema, native DashScope)

import os, dashscope
dashscope.base_http_api_url = 'https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1'
messages = [{"role": "user", "content": [{
    "image": "http://.../receipt.jpg",
    "min_pixels": 32 * 32 * 3,        # 3072
    "max_pixels": 32 * 32 * 8192,     # 8388608
    "enable_rotate": False,           # auto-rotation correction (OCR, native only)
}]}]
response = dashscope.MultiModalConversation.call(
    api_key=os.getenv('DASHSCOPE_API_KEY'),
    model='qwen-vl-ocr-2025-11-20',
    messages=messages,
    ocr_options={
        "task": "key_information_extraction",
        "task_config": {"result_schema": {
            "Ride Date": "ride date, format YYYY-MM-DD",
            "Invoice Code": "invoice code",
            "Invoice Number": "invoice number"
        }}
    },
)
print(response["output"]["choices"][0]["message"].content[0]["text"])

# native HTTP (advanced_recognition with text localization)
curl --location 'https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation' \
  --header "Authorization: Bearer $DASHSCOPE_API_KEY" --header 'Content-Type: application/json' \
  --data '{"model":"qwen-vl-ocr-2025-11-20",
    "input":{"messages":[{"role":"user","content":[
      {"image":"https://.../book.jpg","min_pixels":3072,"max_pixels":8388608,"enable_rotate":false}]}]},
    "parameters":{"ocr_options":{"task":"advanced_recognition"}}}'

---

9. Endpoints / Auth / input limits (vision)

• OpenAI-compatible: https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions (Singapore; ap-southeast-1). Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1 (and https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/... for workspace-scoped).
• Native: .../api/v1/services/aigc/multimodal-generation/generation. Auth: Authorization: Bearer $DASHSCOPE_API_KEY.
• Image: both sides > 10 px; long/short ratio ≤ 200:1; keep ≤ 8K (7680×4320). Formats < 4K: BMP/JPEG/PNG/TIFF/WEBP/HEIC; 4K–8K: only JPEG/JPG/PNG. Size: public URL ≤ 20 MB (Qwen3.7/3.6/3.5) / ≤ 10 MB (others); local ≤ 10 MB; Base64 ≤ 10 MB (source < 7 MB pre-encode).
• Image count: Qwen3-VL plus/flash + 235b-a22b instruct/thinking → image-list 4..2000, URL/path up to 2000; other open-source Qwen3-VL (8b / 30b-a3b / 32b) → 4..512, URL/path up to 256; Base64 up to 250.
• Video: Qwen3-VL up to 2 GB (URL), 2 s..1 h (plus/flash/235b) / up to 20 min (open-source); Base64 < 10 MB; local ≤ 100 MB. Up to 64 videos. Audio-from-video NOT supported (Qwen-Omni only). max_frames (frame cap) — native DashScope SDK only.
• Timeout: non-streaming aborts at 300 s, returns partial (header x-dashscope-partialresponse: true). For long outputs use stream=true.