Hunyuan OCR vs Phi-4-multimodal
Pick Hunyuan OCR for self-hosted photo and scene text spotting, text-image translation, and rare or historical scripts, where a ~1B model returns bounding boxes. Pick Phi-4-multimodal for English document Q&A under an MIT licence, or if you also need audio input. Hunyuan OCR is the only one of the two that outputs coordinates and covers non-English text. Phi-4-multimodal supports vision in English only.
Hunyuan OCRtencent/hunyuan-ocr | Phi-4-multimodalmicrosoft/phi-4-multimodal | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.5 s / page | 2.5 s / page |
| Best for | Chinese documents and photos | Small model for OCR and questions |
| Inputs | PDF, images | images |
| Languages | 100+ languages | English for vision |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | No |
| 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
- It returns bounding boxes plus Markdown, HTML, LaTeX and JSON, which suits jobs that need coordinates or layout.
- One ~1B model covers parsing, text spotting, information extraction and image translation, and it runs on laptops through llama.cpp GGUF.
- It exposes 12 task types, including low-resource and ancient-script OCR, and takes images up to 4K in v1.5.
Choose Phi-4-multimodal if
- The MIT licence is permissive for commercial use, with no region exclusions or user-count threshold.
- It reports DocVQA 93.2 and OCRBench 84.4 at 5.6B parameters, which fits English document Q&A on modest GPUs.
- It takes audio as well as images in one model, and you can self-host it or call it through Azure AI Foundry.
Key differences
| Hunyuan OCR | Phi-4-multimodal | |
|---|---|---|
| Licence | Tencent Hunyuan Community License. Excludes the EU, UK and South Korea. Separate licence needed above 100M monthly active users. | MIT |
| How you run it | Self-host only, via vLLM, Transformers or llama.cpp. | Self-host or Azure AI Foundry. |
| Size | ~1B parameters | 5.6B parameters |
| Output formats | Markdown, HTML, LaTeX, JSON, plain text, bounding boxes | Plain text, Markdown, JSON. No coordinates. |
| Languages | Includes low-resource and ancient-script OCR. | Vision is English only per the model card. Performance drops in other languages. |
| Benchmarks | ParseBench-2 overall 67.11 and ParseBench mean 63.86 (v1.5, independent). | DocVQA 93.2, OCRBench 84.4, ChartQA 81.4, InfoVQA 72.7 (reported by the authors). |
| Hardware and known issues | CUDA 13 and vLLM 0.25.1+ for the unified environment. An archived 1.0 note says Transformers and vLLM outputs differed in accuracy. | Flash attention on A100, A6000 and H100. V100 and older need eager attention. The model card warns of repetitive or inconsistent output in long chats. |
Sources
What each one is
Hunyuan OCR
Small end-to-end OCR model from Tencent covering parsing, extraction and photo translation.
Hunyuan OCR detailsPhi-4-multimodal
Microsoft's 5.6B text, vision and audio model (February 2025). Scores 84.4 on OCRBench per its model card.
Phi-4-multimodal 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": ["microsoft/phi-4-multimodal"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Hunyuan OCR alternativesPhi-4-multimodal alternativesBest OCR models in 2026