StackMap
Subscribe

chandra vs Unlimited-OCR

Datalab's SOTA open OCR model: images/PDFs to structured HTML/Markdown/JSON with layout, tables, forms, checkboxes, handwriting and math, in 90+ languages. Local HF or vLLM inference. — versus — Baidu's open OCR VLM that parses entire multi-page documents in one shot — 'unlimited' long-horizon parsing pushing DeepSeek-OCR further. MIT weights on HF; serve via transformers, vLLM or SGLang.

The curated verdict

Both are open OCR vision-language models. Baidu's model bets on one-shot long-horizon parsing of entire multi-page documents; Chandra processes per-page with stronger layout/table/form structure and a broader language benchmark.

chandraUnlimited-OCR
Stars11k14k
Forks1.2k1.2k
LanguagePythonPython
LicenseApache-2.0MIT
Last activity20 days ago13 days ago
Topicsocrocr, rag
Curated connections45

chandra — the curator's take

Currently the strongest open OCR weights on the olmocr benchmark (85.8, above olmOCR 2 and dots.ocr), with handwriting, filled forms and checkboxes as the real differentiators — plus a serious self-built 90-language benchmark where it averages 72.7% vs Gemini 2.5 Flash's 60.8%. `pip install chandra-ocr`, `chandra_vllm`, done; ~2 pages/s real-world on an H100. The catch is licensing: code is Apache-2.0 but the WEIGHTS are OpenRAIL-M — free for research, personal use and sub-$2M startups, commercial self-hosting needs a Datalab license, and their paid API deliberately stays ahead of the open weights. Pick olmOCR for a fully permissive stack, MinerU when you want a whole parsing pipeline rather than the model itself.

Unlimited-OCR — the curator's take

The current bar if you batch-parse long PDFs: one-shot multi-page parsing keeps cross-page structure (tables spanning pages, running sections) that page-by-page OCR pipelines lose, and first-party vLLM/SGLang recipes plus official Docker images make serving unusually painless for a fresh research model. NOT a lightweight dependency — it's a GPU-hungry VLM; for occasional single pages classic OCR or a hosted API is cheaper. Inference-only repo: no training code, and evaluation beyond the paper's claims is on you.