AI Insights · Oct 22, 2026 · Verified
Small vision models now read PDFs best
For scientific PDFs, small specialised vision-language models now beat both traditional parsing pipelines and large general models. Claude and Gemini also read PDFs directly.
Claim
For bulk parsing of scientific PDFs the current default is a small specialised vision-language model (0.3B to 1B class), not a classic layout pipeline. For a single "read this paper" step, sending the PDF directly to Claude or Gemini is viable and skips the parsing stage entirely.
Evidence
- OmniDocBench v1.6 leaderboard (2026-09-11): TeleOCR 96.91, OvisOCR2 96.47, PaddleOCR-VL-1.6 96.34, MinerU2.5-Pro 95.75, GLM-OCR 95.22, then Gemini 3 Pro 92.91, GPT-5.2 86.59, MinerU-Pipeline 86.47, Mistral OCR 85.66, Marker 78.44. On the Formula column the MinerU and PaddleOCR-VL family scores about 97 versus Marker's 85. Docling is indexed but not on the current board.
- Docling is the mainstream open "document to Markdown/JSON" library (69.4k stars, MIT, LF AI & Data), ships a 258M GraniteDocling VLM pipeline and an MCP server. In LlamaIndex's vendor-run ParseBench (2026-04-13) it scored lowest on semantic formatting, so treat it as a convenient default, not the accuracy leader.
- Native PDF input: Claude accepts 32 MB and up to 600 pages per request with a 1M-token context (100 pages otherwise), processes each page as text plus image, and supports the Files API and citations. Gemini accepts 1000 pages or 50 MB at 258 tokens per page and does not charge for embedded text tokens on Gemini 3.
- Mistral OCR 3 (2025-12-17) claims a 74% win rate over OCR 2 on internal evals but stays mid-pack on the public benchmark. Cheap hosted option, not the accuracy pick.
- ParseBench is vendor-designed and scored; no independent replication found.
Open questions
- Throughput and GPU needs of MinerU2.5-Pro versus PaddleOCR-VL on a CPU-only VM like VM 104.
Sources
- https://github.com/opendatalab/OmniDocBench — repo, Thu Sep 10
- https://github.com/docling-project/docling — repo, Wed Oct 21
- https://platform.claude.com/docs/en/build-with-claude/pdf-support — docs, Wed Oct 21
- https://ai.google.dev/gemini-api/docs/document-processing — docs, Wed Oct 21
- https://www.llamaindex.ai/blog/parsebench — blog, Sun Apr 12
- https://mistral.ai/news/mistral-ocr-3 — blog, Tue Dec 16
AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.