Qwen-OCR text extraction models

POST

/api/v1/services/aigc/multimodal-generation/generation

import os
import dashscope

# Singapore (ap-southeast-1). For another region, replace ap-southeast-1 in the URL.
# API keys are region-specific. See https://docs.qwencloud.com/developer-guides/getting-started/regions
dashscope.base_http_api_url = 'https://ap-southeast-1.maas.qwencloudapi.com/api/v1'

messages = [{
      "role": "user",
      "content": [{
        "image": "https://help-static-aliyun-doc.aliyuncs.com/file-manage-files/zh-CN/20241108/ctdzex/biaozhun.jpg",
        # The minimum pixel threshold for the input image. If the image is smaller than this value, it is scaled up until the total pixels are greater than min_pixels.
        "min_pixels": 32 * 32 * 3,
        # The maximum pixel threshold for the input image. If the image is larger than this value, it is scaled down until the total pixels are less than max_pixels.
        "max_pixels": 32 * 32 * 8192,
        # Specifies whether to enable automatic image rotation.
        "enable_rotate": False}]
      }]
      
response = dashscope.MultiModalConversation.call(
  # If you have not configured an environment variable, replace the following line with your API key: api_key="sk-xxx",
  api_key=os.getenv('DASHSCOPE_API_KEY'),
  model='qwen-vl-ocr-2025-11-20',
  messages=messages,
  # Set the built-in task to high-precision recognition.
  ocr_options={"task": "advanced_recognition"}
)
# The high-precision recognition task returns the result as plain text.
print(response["output"]["choices"][0]["message"].content[0]["text"])

{
  "status_code": 0,
  "request_id": "<string>",
  "code": "<string>",
  "message": "<string>",
  "output": {
    "text": "<string>",
    "finish_reason": "<string>",
    "choices": [
      {
        "finish_reason": "<string>",
        "message": {
          "role": "assistant",
          "content": [
            {
              "text": "<string>",
              "ocr_result": {
                "kv_result": {},
                "words_info": [
                  {
                    "rotate_rect": [
                      0
                    ],
                    "location": [
                      0
                    ],
                    "text": "<string>"
                  }
                ]
              }
            }
          ],
          "logprobs": {
            "content": [
              {
                "token": "<string>",
                "bytes": [
                  0
                ],
                "logprob": 0,
                "top_logprobs": [
                  {
                    "token": "<string>",
                    "bytes": [
                      0
                    ],
                    "logprob": 0
                  }
                ]
              }
            ]
          }
        }
      }
    ]
  },
  "usage": {
    "input_tokens": 0,
    "output_tokens": 0,
    "total_tokens": 0,
    "characters": 0,
    "image_tokens": 0,
    "input_tokens_details": {
      "image_tokens": 0,
      "text_tokens": 0
    },
    "output_tokens_details": {
      "text_tokens": 0
    }
  }
}

Authorizations

string

header

required

API key authentication. Get an API key from the QwenCloud console.

Body

application/json

string

required

The name of the model. See Qwen-OCR for a list of supported models.

Example:qwen-vl-ocr-2025-11-20

object

required

The input object containing messages.

Show child attributes

object[]

required

The context for the model, provided as a sequence of messages.

Show child attributes

enum<string>

required

Must be user.

Available options:user

object[]

required

The message content. Use an array for image input.

Show child attributes

string

The input text. Default: Please output only the text content from the image without any additional descriptions or formatting.

string

The URL, Base64 Data URL, or local path of the image. For more information about passing a local file, see Passing local files.

boolean

defaultfalse

Whether to correct skewed images.

integer

The minimum pixel threshold for the input image in pixels. If an input image has a pixel count below min_pixels, it is enlarged until its total pixel count exceeds min_pixels.

Conversion between image tokens and pixels:

qwen3.5-ocr and qwen-vl-ocr-2025-11-20: Each token corresponds to 32×32 pixels.
qwen-vl-ocr and earlier models: Each token corresponds to 28×28 pixels.

Value range:

qwen3.5-ocr and qwen-vl-ocr-2025-11-20: Default and minimum value is 3072 (3×32×32).
qwen-vl-ocr and earlier models: Default and minimum value is 3136 (4×28×28).

integer

The maximum pixel threshold for the input image in pixels. If the pixel count of the input image is within the [min_pixels, max_pixels] range, the model processes the original image without resizing. If the pixel count exceeds max_pixels, the image is scaled down until its pixel count is less than max_pixels.

Conversion between image tokens and pixels:

qwen3.5-ocr and qwen-vl-ocr-2025-11-20: Each token corresponds to 32×32 pixels.
qwen-vl-ocr and earlier models: Each token corresponds to 28×28 pixels.

Value range:

qwen3.5-ocr and qwen-vl-ocr-2025-11-20: Default 8388608 (8192×32×32), maximum 30720000 (30000×32×32).
qwen-vl-ocr and earlier models: Default 6422528 (8192×28×28), maximum 23520000 (30000×28×28).

object

Model parameters.

Show child attributes

object

Configuration for built-in OCR tasks. When using a built-in task, the model uses a default prompt so you don't need to pass a User Message with text. Minimum SDK versions: Python 1.22.2, Java 2.18.4.

Show child attributes

enum<string>

The built-in task name.

Available options:text_recognition,key_information_extraction,document_parsing,table_parsing,formula_recognition,multi_lan,advanced_recognition

object

Configuration for key_information_extraction task. Specifies fields to extract. If omitted, all fields are extracted.

Show child attributes

object

JSON object specifying fields to extract. Keys are field names, values are descriptions or format requirements. Supports up to 3 levels of nesting.

integer

The maximum number of tokens to generate in the output. If the generated content exceeds this value, the response is truncated.

For qwen3.5-ocr, qwen-vl-ocr-2025-11-20, and qwen-vl-ocr-2024-10-28, the default and maximum values are the same as the model's maximum output length.

integer

A random number seed for reproducible results. Value range: [0, 2^31−1].

number

default0.01

Sampling temperature. Higher = more diverse, lower = more deterministic. Range: [0, 2). Only set one of temperature or top_p.

number

default0.001

Nucleus sampling threshold. Range: (0, 1.0]. Only set one of temperature or top_p.

integer

default1

Candidate set size for sampling. Larger = more random. Must be >= 0.

number

default1

Penalty for repeated sequences. 1.0 means no penalty.

number

default0

Controls content repetition. Range: [-2.0, 2.0].

string

Stop words. Can be a string or array. If stop is an array, you cannot mix token_ids and strings as elements.

boolean

defaultfalse

Whether to return log probabilities. Supported: qwen-vl-ocr-2025-11-20 and later.

integer

default0

Number of most likely tokens to return per step. Range: [0, 5]. Only when logprobs is true.

boolean

defaultfalse

When streaming, true returns only new content per chunk. false returns the full accumulated content.

boolean

defaultfalse

Specifies whether to stream the response.

Python SDK: Set stream=True in the call.
Java SDK: Use the streamCall interface.
HTTP: Set X-DashScope-SSE: enable in the request header.

Response

200-application/json

integer

Request status code. 200 = success. Not returned by Java SDK.

string

Unique identifier for this call. Java SDK returns requestId.

string

Error code. Empty on success. Only returned by Python SDK.

string

Error message. Empty on success.

object

Call result information.

Show child attributes

string | null

Currently fixed to null.

string | null

null during generation, stop when complete, length when truncated.

object[]

Model output.

Show child attributes

string | null

null during generation, stop when complete, length when truncated.

object

Show child attributes

enum<string>

Always assistant.

Available options:assistant

object[]

Output message content.

Show child attributes

string

Text content of the output.

object

Returned for information extraction (key_information_extraction) and high-precision recognition (advanced_recognition) tasks.

Show child attributes

object

Result of the information extraction task.

object[]

Result of the high-precision recognition task.

Show child attributes

number[]

Rotated rectangle [center_x, center_y, width, height, angle]. Angle range: [-90, 90].

number[]

Four-vertex coordinates [x1,y1,x2,y2,x3,y3,x4,y4] in clockwise order from top-left.

string

Text line content.

object | null

Log probability information.

Show child attributes

object[]

Show child attributes

string

integer[]

UTF-8 bytes of the token.

number | null

object[]

Show child attributes

string

integer[]

number | null

object

Token usage information.

Show child attributes

integer

Number of input tokens.

integer

Number of output tokens.

integer

Sum of input_tokens and output_tokens.

integer

Currently fixed to 0.

integer

Tokens for image input.

object

Show child attributes

integer

Tokens for image input.

integer

Tokens for text input.

object

Show child attributes

integer

Tokens in model output.