Hunyuan OCR vs Granite Docling
Pick Granite Docling for English PDFs with tables, code and equations, especially inside a Docling pipeline. It is 258M parameters, Apache-2.0, and its authors report table structure TEDS of 0.97 on FinTabNet. Pick Hunyuan OCR for photo and scene text, image translation, and low-resource or historical scripts, if you are outside the EU, UK and South Korea.
Hunyuan OCRtencent/hunyuan-ocr | Granite Doclingibm/granite-docling | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 1.5 s / page | 1.5 s / page |
| Best for | Chinese documents and photos | Small-model document conversion |
| Inputs | PDF, images | PDF, images |
| Languages | 100+ languages | English, experimental CJK and Arabic |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | No |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose Hunyuan OCR if
- You need text spotting on photos and scenes, or text-image translation, from one ~1B model.
- Your documents use low-resource or ancient scripts, or CJK text, which Granite Docling marks as experimental.
- You want 4K image input and a 128K context window in v1.5, with 12 task types exposed through the vLLM client.
Choose Granite Docling if
- You need a permissive licence: Apache-2.0, with no regional exclusions or user-count limit.
- Your inputs are English PDFs with tables, code and equations: the authors report table TEDS 0.97 on FinTabNet, code F1 0.988 and equation F1 0.968.
- You have a small budget for compute: at 258M parameters it suits CPU or small GPU conversion, and it plugs straight into Docling.
Key differences
| Hunyuan OCR | Granite Docling | |
|---|---|---|
| Licence | Tencent Hunyuan Community License. Excludes the EU, UK and South Korea. A separate licence is needed above 100M monthly active users. | Apache-2.0 |
| Size | ~1B parameters | 258M parameters |
| Hardware | NVIDIA GPU with CUDA 13 recommended for the unified vLLM setup. A llama.cpp GGUF path covers CPU, consumer GPU or laptop. | Small GPU or CPU. On GPUs without bfloat16, such as the T4, it may output only exclamation marks unless run in float32. |
| Output formats | Markdown, HTML, LaTeX, JSON, plain text, bounding boxes | DocTags, Markdown, HTML, LaTeX, bounding boxes |
| Scope | Document parsing, text spotting, information extraction and image translation in one model. | Document conversion to DocTags for the Docling library. |
| Benchmark | ParseBench-2 overall 67.11 and ParseBench mean 63.86 (v1.5, independent). | Table TEDS 0.97 (FinTabNet), full-page OCR F1 0.84, equation F1 0.968, OCRBench 500 (reported by the authors). |
| Languages | Low-resource and ancient-script OCR supported. | English primary. Japanese, Arabic and Chinese are experimental. |
| Known failure | Archived 1.0 note: Transformers and vLLM outputs differed in accuracy. v1.5 needs vLLM 0.25.1+. | Gets stuck in repetitive loops on page headings and invoices. vLLM can fail with a wte attribute error. Ollama on CPU can hallucinate from FP16 overflow. |
Sources
- HunyuanOCR model card
- Hunyuan Community License
- Granite-Docling model card
- HF discussion #24 repetition
- Ollama issue #12355
What each one is
Hunyuan OCR
Small end-to-end OCR model from Tencent covering parsing, extraction and photo translation.
Hunyuan OCR detailsGranite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "tencent/hunyuan-ocr","fallbacks": ["ibm/granite-docling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Hunyuan OCR alternativesGranite Docling alternativesBest OCR models in 2026