baidu/unlimited-ocr

Long-horizon document parsing with batched multi-page OCR. Accepts one image_url or a document_url PDF; no text-only input. The context window is 32,768 tokens, with a maximum output of 24,576 tokens. Aliases include baidu/Unlimited-OCR and unlimited-ocr. The model weights are baidu/Unlimited-OCR.

Output by method

Method	Payload kind	Image `content.object`	Payload
`markdown` (default)	markdown	none	Layout-tagged text with pixel bounding boxes, one `<page>` block per page
`multi_page`	markdown	none	Continuous plain text from an 8-page sliding window, no `<page>` blocks

Both methods return the Markdown-kind payload. multi_page reads an 8-page sliding window per model call by default. Set method_params.window_size to change the window size. The maximum is one image_url or one document_url. Text mode emits a bare payload string for one image, and one <document> block per PDF. With markdown, each document contains one <page> block per page with layout-tagged lines in the form block_type [x1, y1, x2, y2]text. With multi_page, the <document> block carries continuous plain text with no <page> blocks or layout tags. In JSON mode, the response has a data array with kind set to image or document; it does not include top-level model or method keys. See Methods & Response Format.

Request

vlmrun gw chat https://storage.googleapis.com/vlm-data-public-prod/hub/examples/finance.sec-filings/tsla-8k.pdf \
  -m baidu/unlimited-ocr --method markdown

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.vlm.run/v1/openai",
    api_key="<VLMRUN_API_KEY>",
)

response = client.chat.completions.create(
    model="baidu/unlimited-ocr",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "document_url",
                    "document_url": {
                        "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/finance.sec-filings/tsla-8k.pdf"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "markdown"},
)

print(response.choices[0].message.content)

curl https://gateway.vlm.run/v1/openai/chat/completions \
  -X POST \
  -H "Authorization: Bearer $VLMRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "baidu/unlimited-ocr",
    "method": "markdown",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "document_url",
            "document_url": {
              "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/finance.sec-filings/tsla-8k.pdf"
            }
          }
        ]
      }
    ]
  }'

To read a document in sliding windows, use multi_page and set window_size with the Python SDK:

response = client.chat.completions.create(
    model="baidu/unlimited-ocr",
    messages=[...],
    extra_body={
        "method": "multi_page",
        "method_params": {"window_size": 8},
    },
)

The equivalent CLI request is:

vlmrun gw chat tsla-8k.pdf -m unlimited-ocr \
  --method multi_page --method-params '{"window_size": 8}'

Response

Text mode
JSON mode

With markdown, a PDF returns one <document> block with one <page> block per page:

<document filename="tsla-8k.pdf" source_url="…" media_type="application/pdf" pages="5" dpi="96" bytes="38884">
<page index="0" dpi="96" width="816" height="1056">
title [230, 89, 768, 129]UNITED STATES SECURITIES AND EXCHANGE COMMISSION
text [389, 130, 604, 145]WASHINGTON, DC 20549
…
table [21, 803, 978, 842]<table>Title of each classTrading Symbol(s)…</table>
</page>
…
</document>

A single image returns the payload string alone, with no wrapper. With multi_page, the <document> block carries continuous plain text with no <page> blocks or layout tags.

A PDF returns a data array with a document entry and page content:

{
  "data": [
    {
      "kind": "document",
      "source": "tsla-8k.pdf",
      "num_pages": 5,
      "pages": [
        {
          "index": 0,
          "width": 816,
          "height": 1056,
          "content": "title [314, 88, 765, 129]UNITED STATES\nSECURITIES AND EXCHANGE COMMISSION\ntext [389, 130, 603, 144]WASHINGTON, DC 20549\n…"
        }
      ]
    }
  ]
}

With multi_page, the document entry carries one content string and no pages array. An image returns a data array with an image entry:

{
  "data": [
    {
      "kind": "image",
      "image_hash": "sha256:…",
      "image_width": 640,
      "image_height": 480,
      "content": "…"
    }
  ]
}

Overview

Guides

API Reference

Misc

baidu/unlimited-ocr

Output by method

Request

Response

​Output by method

​Request

​Response

Output by method

Request

Response