OCRRouter

GLM-OCR API docs

How OCRRouter calls GLM-OCR, what the upstream returns field by field, and how that becomes Markdown and JSON blocks.

Overview

GLM-OCR is a 0.9B MIT-licensed OCR model from Z.ai that pairs PP-DocLayout-V3 layout analysis with a GLM decoder.

OCRRouter calls it through OpenParser first, which returns layout blocks with pixel boxes. If that fails it falls back to AIHubMix, which returns text only, so the same job can come back with or without layout.

Price and limits

Model ID
zai/glm-ocr
Price
6 credits / page ($6 per 1,000 pages)
Max pages per request
500
Inputs
PDF; PNG, JPEG and other images
Languages
100+ languages
Positions (bbox)
Yes, when the upstream returns them
License
MIT
  • Up to 500 pages per job in OCRRouter
  • OpenParser results carry no confidence for GLM-OCR
  • AIHubMix backup returns no layout in OCRRouter output
  • Model can be changed server-side with OPENPARSER_GLM_MODEL

You pay only for pages that succeed. A failed attempt that falls back to another model is not charged. See pricing.

How OCRRouter calls it

Gateway
OpenParser
Endpoint
POST https://api.openparser.dev/parse (multipart)
Upstream model
ocr_model "glm-ocr"
Request
request part {"ocr_model", "output_format": "openparser@1"} plus the original file; Idempotency-Key header per call
Long documents
HTTP 202 with a job; OCRRouter polls /jobs/{id} with backoff from 1 to 5 seconds
Backup
AIHubMix chat completions, model "glm-ocr", prompt "Text Recognition:", same page rendering as other AIHubMix models
Order
OpenParser first, AIHubMix if it fails. GLM_OCR_UPSTREAM=aihubmix flips the order (server setting)

Options

What OCRRouter sets on the upstream call. You choose the model, format and pages in your request; vendor-specific options are not passed through.

ocr_model
"glm-ocr"
output_format
"openparser@1"
File
Sent as uploaded, no rendering
Client options
None. Callers cannot pass model options.
Backup model / prompt
"glm-ocr" / "Text Recognition:"
Backup temperature
0
Backup PDF render
150 dpi PNG, 4 pages at a time

Raw response fields

Every field the upstream documents, with its type and meaning taken from the vendor's own reference. Fields marked "read by OCRRouter" are the ones our connector uses; the rest are listed so you know what the engine produces.

What OCRRouter receives (OpenParser openparser@1)

OpenParser runs GLM-OCR and converts its result into the openparser@1 document graph: pages with reading order, and elements with locations and scores. Field meanings follow the openparser@1 schema reference.

Source: openparser@1 schema, retrieved 2026-11-19

output_formatstring
Schema version, always "openparser@1".Example: "openparser@1"
document_idstring
Stable identifier for the parsed document.
provenance.provider / provenance.modelstring
Provider and model that produced the result.Example: "baidu" / "paddleocr-vl-1.6"
textstring, read by OCRRouter
Plain text in reading order. Spans index into it.
markdownstring, read by OCRRouter
Best-effort Markdown rendering of the whole document.
pages[].numberinteger, read by OCRRouter
One-based page number, contiguous 1 to N.Example: 1
pages[].width / heightnumber, read by OCRRouter
Size of the page coordinate space.Example: 800 / 300
pages[].unitstring, read by OCRRouter
Unit of the coordinate space: pixel, point, inch or normalized.Example: "pixel"
pages[].element_idsstring[], read by OCRRouter
IDs of elements on the page.
pages[].reading_orderstring[], read by OCRRouter
Ordered top-level element IDs for the page.
elements[].idstring, read by OCRRouter
Element ID, unique in the document.
elements[].kindstring, read by OCRRouter
Element type discriminator: text, table, figure, formula and others.Example: "text"
elements[].rolestring, read by OCRRouter
For text elements: heading, paragraph, line, word, page_header, page_footer and more.Example: "heading"
elements[].textstring, read by OCRRouter
Text of the element.Example: "Total due: $42.50 by December 1."
elements[].locations[].page_numberinteger, read by OCRRouter
Page the location is on.Example: 1
elements[].locations[].bboxobject, read by OCRRouter
Axis-aligned box {left, top, right, bottom} in page coordinates; right and bottom are exclusive.Example: {"left":37,"top":107,"right":401,"bottom":135}
elements[].locations[].polygonPoint[]
Tighter outline when the provider supplies one.
elements[].source.native_id / native_type / native_labelstring
The engine's own ID, type and label for the block. Not in the published schema page; present in responses.Example: "1" / "layout_block" / "text"
elements[].row_count / column_count / cells[]number / array, read by OCRRouter
For table elements: grid size and cells with row_index, column_index, text.
relations[], text_annotations[], assets[]array
Links between elements, span metadata, and page or figure images.Example: []

Real response captured 2026-11-19 from POST /parse on an 800x300 PNG with three lines of text. Trimmed to the first two elements; nothing else edited.

{"output_format": "openparser@1","document_id": "3fd622b1876c999f4f0c3e48a415c085","provenance": {"provider": "zai","model": "glm-ocr","operation": "layout_parsing"},"text": "Invoice 2026\n\nTotal due: $42.50 by December 1.\n…","markdown": "## Invoice 2026\n\nTotal due: $42.50 by December …","pages": [{"number": 1,"source_page_number": 1,"width": 800,"height": 300,"unit": "pixel","rotation_degrees": 0,"languages": [],"element_ids": ["3fd622b1876c999f4f0c3e48a415c085-glm-1-0","3fd622b1876c999f4f0c3e48a415c085-glm-1-1"],"reading_order": ["3fd622b1876c999f4f0c3e48a415c085-glm-1-0","3fd622b1876c999f4f0c3e48a415c085-glm-1-1"]}],"elements": [{"id": "3fd622b1876c999f4f0c3e48a415c085-glm-1-0","locations": [{"page_number": 1,"bbox": { "left": 38, "top": 27, "right": 246, "bottom": 63 }}],"source": {"native_id": "0","native_type": "layout_block","native_label": "paragraph_title"},"kind": "text","role": "heading","text": "Invoice 2026","spans": [],"languages": []},{"id": "3fd622b1876c999f4f0c3e48a415c085-glm-1-1","locations": [{"page_number": 1,"bbox": { "left": 37, "top": 106, "right": 401, "bottom": 135 }}],"source": {"native_id": "1","native_type": "layout_block","native_label": "text"},"kind": "text","role": "paragraph","text": "Total due: $42.50 by December 1.","spans": [],"languages": []}],"text_annotations": [],"relations": [],"assets": []}

What OCRRouter receives from the backup (AIHubMix chat completion)

AIHubMix exposes glm-ocr through its OpenAI-compatible POST /v1/chat/completions endpoint. OCRRouter sends one request per page image and gets back a standard chat completion object. Field meanings follow OpenAI's chat completion object reference, which AIHubMix follows.

Source: OpenAI chat completion object, retrieved 2026-11-19

idstring
Unique identifier for the chat completion.Example: "as-gkcbgipdih"
objectstring
Object type, always "chat.completion".Example: "chat.completion"
createdinteger
Unix timestamp in seconds when the completion was created.Example: 1795104850
modelstring
Model that produced the completion.Example: "deepseek-ocr"
choices[]array, read by OCRRouter
List of completion choices. OCRRouter sends no n parameter, so there is one.
choices[].indexinteger
Position of the choice in the list.Example: 0
choices[].message.rolestring
Role of the message author, "assistant".Example: "assistant"
choices[].message.contentstring or null, read by OCRRouter
Text the model generated. For these OCR models this is the page transcription.Example: "## Invoice 2026\n\nTotal due: ..."
choices[].message.refusalstring or null
Refusal message if the model refused; null otherwise.Example: null
choices[].finish_reasonstring
Why generation stopped: "stop" for a natural end, "length" when the token limit was hit.Example: "stop"
choices[].logprobsobject or null
Log probabilities, only when requested. OCRRouter does not request them.Example: null
choices[].glm_ocr_detail.layout_details[][]array
Per-page layout blocks with bbox_2d, content, label, native_label, index, width, height. Not in AIHubMix docs. OCRRouter ignores it today.
choices[].glm_ocr_detail.data_infoobject
num_pages and per-page width and height.
usage.prompt_tokensinteger
Tokens in the prompt, including the image.Example: 593
usage.completion_tokensinteger
Tokens in the generated output.Example: 81
usage.total_tokensinteger
Prompt plus completion tokens. AIHubMix bills per token.Example: 674

Real response captured 2026-11-19 by OCRRouter with the exact request shown under Options, on an 800x300 PNG with three lines of text. Response id kept; nothing else edited.

{"id": "20261120001419f6803a31f3f44f78","object": "chat.completion","created": 1795104859,"model": "glm-ocr","choices": [{"index": 0,"message": {"role": "assistant","content": "## Invoice 2026\n\nTotal due: $42.50 by December …"},"finish_reason": "stop","glm_ocr_detail": {"layout_details": [[{"bbox_2d": [38, 28, 247, 64],"content": "## Invoice 2026","height": 300,"index": 0,"label": "text","native_label": "paragraph_title","width": 800},{"bbox_2d": [37, 107, 401, 135],"content": "Total due: $42.50 by December 1.","height": 300,"index": 1,"label": "text","native_label": "text","width": 800},{"bbox_2d": [37, 158, 347, 187],"content": "Thank you for your business.","height": 300,"index": 2,"label": "text","native_label": "text","width": 800}]],"layout_visualization": [],"data_info": {"num_pages": 1,"pages": [{ "height": 300, "width": 800 }]}}}],"usage": {"prompt_tokens": 118,"completion_tokens": 30,"total_tokens": 148}}

GLM-OCR native output (Z.ai layout_parsing API)

What Z.ai's own POST /paas/v4/layout_parsing returns. OpenParser maps layout_details into elements. Note the docs say bbox_2d is normalized 0 to 1, but the AIHubMix response above carries pixel values.

Source: Z.ai layout parsing API, retrieved 2026-11-19

idstring
Task ID.
createdinteger
Creation time, Unix seconds.
modelstring
Model that handled the request.
md_resultsstring
Recognized content as Markdown.
layout_details[][]array of arrays
Layout blocks, one inner array per page.
layout_details[][].indexinteger
Element index.
layout_details[][].labelstring
Element type: image, text, formula or table.
layout_details[][].bbox_2dnumber[4]
Element box [x1, y1, x2, y2], documented as normalized 0 to 1.
layout_details[][].contentstring
Text, image URL or table HTML depending on label.
layout_details[][].width / heightinteger
Page width and height.
layout_visualizationstring[]
URLs of result images.
data_info.num_pagesinteger
Total pages.
data_info.pages[]array
Width and height per page.
usage.prompt_tokens / completion_tokens / total_tokensinteger
Token counts.
usage.prompt_tokens_details.cached_tokensinteger
Tokens served from cache.
request_idstring
Request ID.

Built from the example values in Z.ai's API reference, not a captured response.

{"id": "task_123456789","created": 1727156815,"model": "GLM-OCR","md_results": "# Doc title\nThis is the document content...","layout_details": [[{"index": 1,"label": "text","bbox_2d": [0.1, 0.1, 0.5, 0.3],"content": "This is the content of the element","width": 600,"height": 800}]],"layout_visualization": ["https://example.com/visualization/page1.png"],"data_info": {"num_pages": 5,"pages": [{ "width": 600, "height": 800 }]},"usage": {"prompt_tokens": 1024,"completion_tokens": 512,"prompt_tokens_details": { "cached_tokens": 0 },"total_tokens": 1536},"request_id": "req_123456789"}

How it maps to OCRRouter output

pages[] sorted by number
One OCRRouter page each
pages[].reading_order (else element_ids)
Order of layout blocks and Markdown on the page
elements[].kind + role
Block type: table, formula, figure/image/picture -> image, heading/title -> heading, else paragraph
elements[].text (or markdown)
Block text. Headings are normalized to one run of #, level 1 to 6
Table cells
Rebuilt as a GitHub-flavoured Markdown table; first row is the header
locations[].bbox
Block bbox [left, top, right, bottom], using the location on the same page
pages[].width, height, unit
Page layout width, height and unit (pixel here)
confidence.score
Block confidence (0 to 1); null when absent
Page Markdown
Block texts joined with blank lines
Document markdown
Fallback only when no element has text: all of it goes on page 1
Backup: choices[0].message.content
Page text, HTML tables rewritten as Markdown; no layout, bbox or confidence

The unified output is described in Response format.

Example request

curl https://api.ocrrouter.ai/v1/ocr \
  -H "Authorization: Bearer $OCRROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai/glm-ocr",
    "file_url": "https://example.com/invoice.pdf",
    "format": "json"
  }'

Sources

Vendor documentation read on 2026-11-19.