Docling vs InternVL3.5
Pick Docling to ingest born-digital PDFs and Office files into a RAG pipeline, because it parses many formats and exports lossless JSON with layout and reading order. Pick InternVL3.5 for document question answering on a GPU, because it is a vision language model you can run at sizes from 1B to 241B-A28B. Docling gives you structure but needs an OCR engine for scans. InternVL3.5 gives you answers but no bounding boxes or layout output by default.
Doclingibm/docling | 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 | 2.2 s / page | 3.0 s / page |
| Best for | Structured export for RAG | Self-hosted document understanding |
| Inputs | PDF, images, DOCX | images |
| Languages | Latin scripts, CJK | 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 | No | ✓ 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 Docling if
- You need one API for PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, emails and XBRL.
- You need structured JSON that keeps layout, reading order and table structure for downstream tools.
- You want to run on CPU on macOS, Linux or Windows, and plug in EasyOCR, Tesseract or RapidOCR.
Choose InternVL3.5 if
- You want document question answering from a single model rather than a conversion pipeline.
- You want to pick a size from nine options, from 1B to 241B, to trade accuracy for cost.
- You have GPUs: models up to 30B run on a single A100, and the MoE variants 20B-A4B and 30B-A3B activate few parameters per token.
Key differences
| Docling | InternVL3.5 | |
|---|---|---|
| What it is | Python library from IBM Research that converts documents into a structured DoclingDocument | Open-weight vision language model family from OpenGVLab, 1B to 241B-A28B parameters |
| Licence | MIT (individual models have their own licences) | Apache-2.0 |
| How you run it | Local library; CPU works, GPU speeds up model stages | Self-host with Transformers, LMDeploy or vLLM; up to 30B on one A100, 38B on two, 241B-A28B on eight |
| Price | Free, self-hosted (you pay for compute) | Free, self-hosted (you pay for compute) |
| Output formats | Markdown, HTML, JSON, DocTags, plain text | Plain text, Markdown, JSON; no bounding boxes or layout output by default |
| Scanned pages | Accuracy depends on the OCR engine you plug in; GraniteDocling VLM option is 258M | OCR scores are published only as chart images, not text values |
| Known failure | Full-page Tesseract OCR in 2.143.1 drops words and numbers (issue #5254); reading order can merge neighbouring columns (issue #5132) | Larger sizes need multiple A100s and do not suit CPU-only deployment |
Sources
- Docling GitHub README
- Supported formats
- Issue #5254 OCR regression
- Issue #5132 reading order
- Hugging Face model card (8B)
- InternVL GitHub
What each one is
InternVL3.5
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": "ibm/docling","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Docling alternativesInternVL3.5 alternativesBest OCR models in 2026