GOT-OCR 2.0 vs Phi-4-multimodal
Pick GOT-OCR 2.0 for formulas, sheet music, chemical structures and region-specific OCR, where it outputs LaTeX, TikZ, SMILES and kern. Pick Phi-4-multimodal for English document Q&A and mixed audio and image pipelines, mainly because it has a published DocVQA score of 93.2, an MIT licence and an Azure option. Neither covers non-English well: GOT-OCR 2.0 is aimed at English and Chinese, and Phi-4-multimodal supports vision in English only.
GOT-OCR 2.0stepfun/got-ocr-2 | Phi-4-multimodalmicrosoft/phi-4-multimodal | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.5 s / page | 2.5 s / page |
| Best for | Formatted text, formulas and charts | Small model for OCR and questions |
| Inputs | images | images |
| Languages | Chinese, English | English for vision |
| Max pages per request | 100 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | 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 GOT-OCR 2.0 if
- It outputs LaTeX, TikZ, SMILES and kern, so it covers formulas, sheet music and chemical structures.
- It supports region OCR guided by box coordinates or color, so you can read one part of a page.
- It is a 580M-parameter model that reads English and Chinese documents, and it is merged into Hugging Face Transformers with batched inference.
Choose Phi-4-multimodal if
- Microsoft reports DocVQA 93.2 and OCRBench 84.4, so you have published numbers to check against your documents.
- The MIT licence is permissive for commercial use, and you can self-host or call it through Azure AI Foundry.
- It takes audio as well as images, with a 128K token context, so one model can serve a mixed audio and image pipeline.
Key differences
| GOT-OCR 2.0 | Phi-4-multimodal | |
|---|---|---|
| Licence | Code Apache-2.0; training data CC BY-NC 4.0 | MIT |
| How you run it | Self-host only. CUDA 11.8 and PyTorch 2.0.1 per README; CPU users are pointed to a llama.cpp port. | Self-host or Azure AI Foundry. Flash attention on A100, A6000 and H100; V100 and older need eager attention. |
| Size | 580M parameters | 5.6B parameters |
| Output formats | Plain text, Markdown, LaTeX, TikZ, SMILES, kern | Plain text, Markdown, JSON |
| Published benchmarks | Not published. The README lists no scores. | DocVQA 93.2, OCRBench 84.4, ChartQA 81.4, InfoVQA 72.7 (reported by Microsoft) |
| Languages | English and Chinese documents | Vision supported for English only; performance drops in other languages |
| Known failure | Output can loop characters midway (issue #226), return wrong text on blank images (issue #235), and format mode is much worse than plain OCR mode (issue #317). | The model card warns of repetitive or inconsistent output in long chats. |
Sources
What each one is
GOT-OCR 2.0
Unified OCR model that reads plain text, tables, formulas, sheet music and charts.
GOT-OCR 2.0 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": "stepfun/got-ocr-2","fallbacks": ["microsoft/phi-4-multimodal"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
GOT-OCR 2.0 alternativesPhi-4-multimodal alternativesBest OCR models in 2026