DeepSeek-OCR vs InternVL3.5
Pick DeepSeek-OCR for cheap self-hosted bulk OCR of pages where you want Markdown, HTML or bounding boxes back, and you can add safeguards against repeated tokens. Pick InternVL3.5 for document question answering or a multimodal pipeline where you want one model family in several sizes and JSON output. The main split is that DeepSeek-OCR is a dedicated OCR model with layout output, while InternVL3.5 is a general vision language model without bounding boxes by default.
DeepSeek-OCRdeepseek/deepseek-ocr | InternVL3.5opengvlab/internvl | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.1 s / page | 3.0 s / page |
| Best for | Cheap bulk conversion | Self-hosted document understanding |
| Inputs | PDF, images | images |
| Languages | 100+ languages | Multilingual |
| Max pages per request | 200 | 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 DeepSeek-OCR if
- It returns bounding boxes and HTML as well as Markdown and plain text, so you get layout from the model itself.
- It offers resolution modes from 64 to 400 vision tokens per page, and its authors quote about 2,500 tokens per second of PDF throughput on one A100-40G.
- It uses the MIT license, which allows commercial use, and it is built for cheap self-hosted bulk OCR.
Choose InternVL3.5 if
- It comes in nine sizes from 1B to 241B, so you can trade accuracy for cost within one model family.
- It can output JSON and suits document question answering and self-hosted multimodal pipelines.
- Models up to 30B fit on a single A100, and the 20B-A4B and 30B-A3B MoE variants activate few parameters per token.
Key differences
| DeepSeek-OCR | InternVL3.5 | |
|---|---|---|
| Model type | Open-weight OCR model that compresses pages into few vision tokens | General open-weight vision language model family from OpenGVLab |
| License | MIT | Apache-2.0 |
| How you run it | Self-host with Transformers or vLLM on a CUDA GPU (reference setup CUDA 11.8, PyTorch 2.6, bfloat16) | Self-host with Transformers, LMDeploy or vLLM; up to 30B on one A100, 38B on two A100s, 241B-A28B on eight A100s |
| Price | Free, you pay for compute | Free, you pay for compute |
| Output formats | Markdown, plain text, HTML, bounding boxes | Plain text, Markdown, JSON; no bounding boxes or layout output by default |
| Sizes | One model with resolution modes of 64 to 400 vision tokens per page | Nine sizes from 1B to 241B-A28B (8B variant is 8.5B parameters) |
| Known failure | Repeated tokens and looping over table headers (GitHub issues #12 and #16); it corrects nonsense words instead of copying them | OCR scores are published only as chart images, not text values |
Sources
- DeepSeek-OCR GitHub
- Visual Merit or Linguistic Crutch paper
- LearnOpenCV test
- 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": "deepseek/deepseek-ocr","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
DeepSeek-OCR alternativesInternVL3.5 alternativesBest OCR models in 2026