Best open OCR and document models
OCR and document models extract text from images, PDFs, and scanned documents. They can also understand document layout, tables, and structure. These models are useful for digitizing documents, parsing receipts, and extracting data from forms.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes |
|---|---|---|---|---|---|---|
| 1 | Unlimited-OCR baidu | – · FP16 | – | MIT | 4.2M | 4.9K |
| 2 | DeepSeek-OCR deepseek-ai | – · FP16 | – | MIT | 2M | 3.5K |
| 3 | GLM-OCR zai-org | – · FP16 | – | MIT | 1.7M | 2.2K |
| 4 | manga-ocr-base kha-white | – · FP16 | – | Apache 2.0 | 601.8K | 192 |
| 5 | PP-DocLayoutV3_safetensors PaddlePaddle | – · FP16 | – | Apache 2.0 | 930.3K | 58 |
| 6 | trocr-base-printed microsoft | – · FP16 | – | Not stated | 273.8K | 234 |
| 7 | PP-OCRv5_server_det PaddlePaddle | – · OTHER | – | Apache 2.0 | 428.2K | 75 |
| 8 | Unlimited-OCR-2 baidu | – · FP16 | – | MIT | 147.6K | 564 |
| 9 | trocr-large-printed microsoft | – · FP16 | – | Not stated | 168K | 188 |
| 10 | trocr-base-handwritten microsoft | – · FP16 | – | MIT | 99K | 534 |
| 11 | trocr-base-handwritten-hist-swe-2 Riksarkivet | – · FP16 | – | Apache 2.0 | 240.3K | 18 |
| 12 | trocr-small-handwritten microsoft | – · FP16 | – | Not stated | 111.7K | 75 |
| 13 | PP-OCRv5_server_rec PaddlePaddle | – · OTHER | – | Apache 2.0 | 154K | 32 |
| 14 | trocr-large-handwritten microsoft | – · FP16 | – | Not stated | 39.2K | 174 |
| 15 | PP-OCRv5_mobile_det PaddlePaddle | – · OTHER | – | Apache 2.0 | 84.4K | 31 |
How we rank
Models whose name or metadata signals OCR or document understanding capability, excluding text-to-image models. We include models seen on Hugging Face in the last 14 days with at least 10,000 downloads in the last 30 days. Score = log10(30-day downloads) + ½ × log10(likes + 1), so popularity leads and likes break near-ties. We keep one entry per model name and leave out adult and uncensored builds. Memory is estimated from parameter count and the published file format (16-bit ≈ 2 bytes per parameter, 4-bit ≈ 0.5) plus 20% headroom; long contexts need more. A quantised copy of a bigger model may also fit — check the model page.