Document OCR for Skewed, Curved and Camera-Captured Pages
Reads a page image and returns its text with layout and structure: tables, formulas and reading order. It works on skewed or curved pages without a separate dewarping step. Document-parsing VLM from XingChen-AGI (TeleAI), Apache-2.0. Strongest on Chinese and English.
What it does
TeleOCR is a compact vision-language model built only for document parsing. Give it a page image and an instruction, and it transcribes the page: body text, tables, formulas and layout.
What sets it apart is distorted pages. Phone photos of a book spread, a curled receipt or a page shot at an angle usually need flattening before OCR. TeleOCR reads them as they are. The model card shows results on the DocUNet and DIR300 warped-document datasets.
It is an OCR model, not a chat assistant. Ask it to read an image and it reads it exactly. Ask it a conversational question about a picture and you get transcription-shaped text back. For questions about an image, use a multimodal language model like Gemma 4 or Qwen3.8 instead.
Languages: trained and benchmarked on Chinese and English.
Problem it solves
Phone photos of documents | Skewed, curved or angled pages read directly, no flattening step |
Tables | Recovers table structure instead of a flat run of text |
Formulas | Returned as LaTeX from papers, textbooks and worksheets |
Code snippets | Reads code printed in documents and screenshots |
Deterministic output | Same page in, same text out at the default settings, which matters for batch runs |
Input/Output
- Input: Two ports
- Image — a document page. Format: JPG, PNG, WEBP, AVIF. Content: printed pages, scans, or camera photos of documents
- Text — the instruction, e.g. "Read all the text in this image." Keep it a reading instruction, not a question
- Output: Text — the page transcription with its structure
Settings (advanced)
- max_new_tokens — default 1024, covers an ordinary page. If a dense page stops mid-sentence, raise this (up to 8192)
- temperature — default 0, deterministic. Leave it at 0. Raising it makes the model guess plausible characters where the scan is unclear
- top_p / top_k — only apply when temperature is above 0
Choosing an OCR model
Your input | Use |
Chinese or English pages photographed at an angle, curved or warped | TeleOCR |
Japanese documents — vertical text, tables, formulas | Sarashina2.2-OCR |
Korean, Chinese or English printed text, high volume | PaddleOCR |
Full-page layout understanding, general multi-language documents | GLM-OCR |
Handwriting | DeepSeek OCR |
Performance
- Benchmarks (reported by the developer): OmniDocBench v1.6 overall 96.87 (table TEDS 97.05, formula CDM 96.36); Wild_OmniDocBench overall 88.53, the top score in the model card's comparison for in-the-wild documents
- Strongest on: Chinese and English documents with real structure, including camera-captured and distorted pages
- Watch for: long, dense pages getting cut off — raise max_new_tokens rather than re-running
- Not for: Japanese vertical text (use Sarashina2.2-OCR), Korean-heavy volume work (use PaddleOCR), handwriting (use DeepSeek OCR), or open questions about a picture (use a multimodal language model)
Technical Details
Architecture | Compact document-parsing VLM built on Qwen2.5-VL (about 1.2B parameters per the model card) |
Developer | XingChen-AGI (TeleAI) |
Languages | Chinese, English |
Structure recovery | Text · tables · formulas as LaTeX · code · layout, including distorted pages without dewarping |
Runtime | GPU, cloud — no local setup |
Compliance & Provenance
Provider | XingChen-AGI (TeleAI) |
Provider type | Specialized |
License | Apache-2.0 |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | OCR output is labeled "AI-extracted", not "AI-generated" |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- Model card
- License: Apache-2.0