Qianfan-OCR
Qianfan-OCR is an end-to-end document model. Its card describes a 4B model for document recognition, layout understanding, questions about documents and key-field extraction across 192 languages.
Model design and access
Layout-as-Thought produces region types, locations and reading order before the final content. It is optional. The Qianfan vision encoder supports dynamic image resolution up to 4K, with image tiles used to preserve document detail.Qianfan-OCR official model card
Practical use
Practical guidance. Enable the layout stage for pages with mixed columns, figures and tables. For a uniform form, compare the extra processing with direct extraction. Require exact field names and inspect empty or ambiguous values so a fluent response does not conceal missing source text.Qianfan-OCR official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| License | Apache 2.0Qianfan-OCR official model card |
| Output | Markdown and task-specific textQianfan-OCR official model card |
| Layout option | Layout-as-ThoughtQianfan-OCR official model card |
Related models
- Unlimited-OCR-2
Baidu · Unlimited-OCR
Document parsing model with sparse-reference sliding-window attention for long documents.
Launch date Downloadable · API - PaddleOCR-VL-1.6
Baidu · PaddleOCR-VL
A 0.9-billion-parameter document vision-language model with improved training for difficult document regions.
Launch date Downloadable