IndicOCR
IndicOCR combines a trained document-layout model with a trained text-recognition model. The joint release produces reading-ordered Markdown and per-block JSON from page images.
Restricted access
Layout detection and recognition are separate stages
The layout model uses 37 education-oriented classes. Each stage can run separately through a JSON interface. Some detected blocks are intentionally not transcribed; an empty text field does not necessarily mean detection failed. The card identifies complex reading order and handwriting variation as continuing limitations.IndicOCR official model card
Inspect block order as well as recognized words
Practical guidance. A page can contain correct words in the wrong sequence. For textbooks and exam papers, compare the layout JSON with the source page before using the extracted text for lessons or search.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable, API |
| Model identifiers | indic-ocrIndicDocLayoutIndicBlockOCR |
| Trained components | IndicDocLayout: 33M parameters, adapted from PP-DocLayoutV3/RT-DETR. IndicBlockOCR: 0.8B parameters, based on Qwen3.5-0.8B.IndicOCR official model card |
| Tokenizer | IndicBlockOCR uses the Sarvam-30B tokenizer.IndicOCR official model card |
| Language coverage | Printed pages: English and 22 Indian languages. Handwriting: English and 12 Indian languages.IndicOCR official model card |
| Access and terms | Gated weights under Indic Open Model License v1.0; upstream terms also apply.Indic Open Model License v1.0 |
| Hosted access | Bodhan documents an authenticated API operation for this model.Bodhan API reference |
Related models
- Indic Translate
Bodhan AI · Indic Translate
7.94-billion-parameter Gemma-derived model for document translation between English and 22 Indian languages.
Launch date not confirmedRestricted accessDownloadable · API