dots.ocr vs InternVL3.5
Pick dots.ocr when you need structured output from PDFs: layout boxes, tables as HTML and formulas as LaTeX from one 3B model. Pick InternVL3.5 when you need document question answering or one model family at several sizes. The main split is that dots.ocr is a dedicated parser with coordinates, while InternVL3.5 is a general VLM with no bounding boxes by default.
dots.ocrrednote/dots-ocr | InternVL3.5opengvlab/internvl | |
|---|---|---|
| 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 | 3.0 s / page |
| Best for | Multilingual document parsing | Self-hosted document understanding |
| Inputs | PDF, images | images |
| Languages | 100+ languages | Multilingual |
| 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 | ✓ Yes |
| 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 dots.ocr if
- It returns layout boxes, categories and text in one pass, with tables as HTML and formulas as LaTeX.
- The authors report 79.1 on olmOCR-bench for the original model and 83.9 for dots.mocr (v1.5), so there is a published OCR score to check.
- At 3B parameters (1.7B LLM backbone) it runs on a single CUDA GPU with vLLM, and a CPU inference guide is linked.
Choose InternVL3.5 if
- Nine sizes from 1B to 241B-A28B let you trade accuracy for cost, and the MoE variants 20B-A4B and 30B-A3B activate few parameters per token.
- It suits document question answering, not only text extraction.
- The Apache-2.0 licence has no extra agreement file, and models up to 30B fit on a single A100 per the model card.
Key differences
| dots.ocr | InternVL3.5 | |
|---|---|---|
| Licence | MIT, with a separate dots.ocr license agreement file in the repository | Apache-2.0 |
| Model size | 3B (1.7B LLM backbone) | 1B to 241B-A28B; the 8B variant is 8.5B |
| Output formats | JSON, Markdown, HTML, LaTeX, bounding boxes, SVG | Plain text, Markdown, JSON |
| Bounding boxes and layout | Returns layout boxes and categories | None by default, so not suited to jobs needing word coordinates |
| OCR benchmark | olmOCR-bench 79.1 (dots.ocr) and 83.9 (dots.mocr), reported by the authors | Not published as text; the model card shows OCR scores only as chart images |
| Hardware | Single CUDA GPU with vLLM is the documented path; no VRAM figure stated | Up to 30B on one A100; 38B needs two A100s; 241B-A28B needs eight A100s |
| Known failure | Runs of ellipses or underscores can make output repeat endlessly; tables above 1000 grid cells raise an error and drop the page in Docling | Larger sizes need multiple GPUs and are poor fits for CPU-only deployment |
Sources
- dots.ocr GitHub README
- dots.ocr blog / limitations
- Docling issue #5146
- Hugging Face model card (8B)
- InternVL GitHub
What each one is
dots.ocr
Single vision model from rednote that does layout detection and content recognition together.
dots.ocr detailsInternVL3.5
Shanghai AI Lab's open vision-language family (1B to 241B, August 2025), strong on OCR and document benchmarks.
InternVL3.5 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": "rednote/dots-ocr","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
dots.ocr alternativesInternVL3.5 alternativesBest OCR models in 2026