Overview
DeepSeek-OCR is an open-weight MIT-licensed vision-language OCR model from DeepSeek that compresses each page into a small number of vision tokens.
OCRRouter calls it through AIHubMix with the grounding prompt, so each block of output carries a type label and a box. That gives you page Markdown plus layout blocks with positions.
Price and limits
- Model ID
- deepseek/deepseek-ocr
- Price
- 1 credit / page ($1 per 1,000 pages)
- Max pages per request
- 200
- Inputs
- PDF (rendered to page images); PNG, JPEG and other images AIHubMix accepts as image_url
- Languages
- 100+ languages
- Positions (bbox)
- Yes, when the upstream returns them
- License
- MIT
- Up to 200 pages per job in OCRRouter
- One image per request; long PDFs take one call per page
- No confidence scores
- Known to repeat tokens on some tables (GitHub issues #12 and #16)
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
- AIHubMix, OpenAI-compatible
- Endpoint
- POST https://aihubmix.com/v1/chat/completions
- Upstream model
- "deepseek-ocr"
- Request
- One user message per page: the page as a base64 data URL in an image_url part, then the prompt as a text part
- PDFs
- Rendered to PNG at 150 dpi with poppler pdftoppm, one request per page, 4 pages at a time
- Images
- Sent as is, one request
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.
- model
- "deepseek-ocr"
- prompt
- "<|grounding|>Convert the document to markdown."
- temperature
- 0
- PDF render
- 150 dpi PNG
- Concurrency
- 4 pages at a time
- Other parameters
- None sent (no max_tokens, no stream, no response_format)
- Client options
- None. Callers cannot pass model options; OCRRouter fixes the request above.
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 (AIHubMix chat completion)
AIHubMix exposes deepseek-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 from OpenAI's chat completion object reference and our test call on 2026-11-19; the field list below is every field in that response.
Source: OpenAI chat completion object, retrieved 2026-11-19
idstring- Unique identifier for the chat completion.Example:
"as-ryvg77zszb" objectstring- Object type, always "chat.completion".Example:
"chat.completion" createdinteger- Unix timestamp in seconds when the completion was created.Example:
1795106679 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:
"<|ref|>sub_title<|/ref|><|det|>[[49, 64, 419, 163]]<|/det|>\n..." 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 usage.prompt_tokensinteger- Tokens in the prompt, including the image.Example:
493 usage.completion_tokensinteger- Tokens in the generated output.Example:
111 usage.total_tokensinteger- Prompt plus completion tokens. AIHubMix bills per token.Example:
604
Real response captured 2026-11-19 by OCRRouter with the exact request shown under Options, on an 800x420 PNG with a heading, a paragraph and a 3x3 table. Nothing edited.
{"id": "as-ryvg77zszb","model": "deepseek-ocr","object": "chat.completion","created": 1795106679,"choices": [{"index": 0,"finish_reason": "stop","logprobs": null,"message": {"role": "assistant","content": "<|ref|>sub_title<|/ref|><|det|>[[49, 64, 419, 1…","refusal": null}}],"usage": {"prompt_tokens": 493,"completion_tokens": 111,"total_tokens": 604}}DeepSeek-OCR native output (grounding mode)
The model returns one string. With <|grounding|> in the prompt (the default PROMPT in DeepSeek-OCR-vllm/config.py) it writes a <|ref|> label and a <|det|> box before each block. Field meanings from DeepSeek-OCR-vllm/run_dpsk_ocr_image.py and our test call on 2026-11-19: re_match() finds each <|ref|>label<|/ref|><|det|>boxes<|/det|> pair, draw_bounding_boxes() scales each box by x / 999 * image width and y / 999 * image height, image blocks are cropped and replaced with a Markdown image link, and all other markers are stripped to leave the Markdown. The code does not list every label; title, image and the labels below are the ones seen in the code or our test.
Source: DeepSeek-OCR-vllm/run_dpsk_ocr_image.py, retrieved 2026-11-19
<|ref|>...<|/ref|>token span, read by OCRRouter- Block label. The reference code treats image (cropped to a picture) and title (drawn with a thicker box) specially. Our test returned sub_title, text and table.Example:
<|ref|>table<|/ref|> <|det|>[[x0,y0,x1,y1]]<|/det|>token span, list of boxes, read by OCRRouter- One or more boxes as top-left and bottom-right corners, normalised to 0 to 999 of the image width and height. The reference code parses the span as a Python list of boxes and multiplies by size / 999 to get pixels.Example:
<|det|>[[49, 377, 728, 744]]<|/det|> block textstring, read by OCRRouter- Markdown for the block, after the det span. Headings use #, tables come as HTML <table> with no header cells.Example:
<table><tr><td>Region</td><td>Q2</td>...
The message.content string from the real response above, captured 2026-11-19.
<|ref|>sub_title<|/ref|><|det|>[[49, 64, 419, 163]]<|/det|>
# Quarterly Report
<|ref|>text<|/ref|><|det|>[[49, 230, 555, 285]]<|/det|>
Sales grew 12 percent in the third quarter.
<|ref|>table<|/ref|><|det|>[[49, 377, 728, 744]]<|/det|>
<table><tr><td>Region</td><td>Q2</td><td>Q3</td></tr><tr><td>North</td><td>120</td><td>134</td></tr><tr><td>South</td><td>98</td><td>110</td></tr></table>How it maps to OCRRouter output
- choices[0].message.content
- Page text. A wrapping ```markdown fence is stripped.
- HTML <table> in content
- Rewritten as a GitHub-flavoured Markdown table (htmlTablesToMarkdown); the first row is the header
- One response per page
- pages[i] in OCRRouter output, in page order
- <|ref|>label<|/ref|><|det|>[[x0,y0,x1,y1]]<|/det|>
- Starts a layout block. Text up to the next marker becomes the block text
- ref label
- Block type: contains table -> table, title -> heading, image or figure -> image, formula or equation -> formula, else paragraph
- det box
- Block bbox [x0, y0, x1, y1] on a 0 to 999 grid; page layout unit is normalized_1000 with width and height 1000
- Heading blocks
- Get a leading # if the model gave none; level is the count of #
- Confidence
- null. DeepSeek-OCR gives no scores
- Page Markdown
- Block texts joined with blank lines, markers removed. If no markers appear, the whole content is the page and there is no layout
- Layout for the job
- Returned only when every page had markers
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": "deepseek/deepseek-ocr",
"file_url": "https://example.com/invoice.pdf",
"format": "json"
}'Sources
Vendor documentation read on 2026-11-19.