InternVL3.5 vs DeepSeek-VL2
Pick InternVL3.5 if you want one Apache-2.0 model family you can scale from 1B to 241B-A28B and run in Transformers, LMDeploy or vLLM. Pick DeepSeek-VL2 for single-page document and chart question answering, where it has published scores: DocVQA 93.3 and ChartQA 86.0. The main split is licence and evidence: InternVL3.5 is the more permissive licence with more size options, while DeepSeek-VL2 has OCR numbers you can check.
InternVL3.5opengvlab/internvl | DeepSeek-VL2deepseek/deepseek-vl2 | |
|---|---|---|
| Type | Vision model | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 3.0 s / page | 3.0 s / page |
| Best for | Self-hosted document understanding | Charts and tables |
| Inputs | images | images |
| Languages | Multilingual | Mainly English and Chinese |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| Bounding boxes | No | ✓ Yes |
| 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 InternVL3.5 if
- It is Apache-2.0, built on Apache-2.0 Qwen3, so there is no custom weights licence to review.
- It comes in nine sizes from 1B to 241B, and models up to 30B fit on a single A100.
- It outputs plain text, Markdown and JSON, which suits document question answering and self-hosted multimodal pipelines.
Choose DeepSeek-VL2 if
- It has published OCR and document scores: DocVQA 93.3 for the full model, OCRBench 834 for Small and ChartQA 86.0.
- Its mixture-of-experts design activates only 1.0B to 4.5B parameters per token.
- The weights allow commercial use, and it targets document and chart question answering on single pages.
Key differences
| InternVL3.5 | DeepSeek-VL2 | |
|---|---|---|
| Licence | Apache-2.0 | MIT for code; DeepSeek Model License for weights (commercial use allowed) |
| Sizes | Nine sizes, 1B to 241B-A28B (8B variant is 8.5B) | Three sizes: Tiny 3.37B (1.0B active), Small 16.1B, full 27.5B (4.5B active) |
| GPU needs | Up to 30B on a single A100; 38B needs two A100s; 241B-A28B needs eight A100s | Small may need 80GB, or 40GB with incremental prefilling; Tiny runs under 40GB |
| Published OCR scores | Not published as text; the model card shows OCR scores only as chart images | DocVQA 93.3 (full), OCRBench 834 (Small), ChartQA 86.0 (full), from the authors' paper |
| Output formats | Plain text, Markdown, JSON | Plain text, Markdown |
| Multi-image input | Not published | Only a few images per session; dynamic tiling turns off with more than two images |
| Known failure | No bounding boxes or layout output by default, so no word coordinates | Struggles with blurry images, so phone photos are a risk |
Sources
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 detailsDeepSeek-VL2
DeepSeek's mixture-of-experts vision models (December 2024) with OCR and table and chart understanding.
DeepSeek-VL2 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": "opengvlab/internvl","fallbacks": ["deepseek/deepseek-vl2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
InternVL3.5 alternativesDeepSeek-VL2 alternativesBest OCR models in 2026