🪴 Brain

Unlimited-OCR

2 min read

Unlimited-OCR

End-to-end OCR model from Baidu that parses dozens of pages of documents in a single forward pass under a standard 32K max length — “one-shot long-horizon parsing.” Builds directly on DeepSeek-OCR, pushing it one step further. 3B params, BF16, Apache-2.0.

Description

The key idea: end-to-end OCR with an LLM decoder benefits from the language prior, but as output lengthens the accumulated KV cache balloons memory and slows generation — unlike a human, who copies long documents without efficiency decline. Unlimited-OCR emulates human “parsing working memory” by replacing all decoder attention layers with Reference Sliding Window Attention (R-SWA), which keeps a constant KV cache across the whole decode and cuts attention compute. Combined with DeepSeek-OCR’s high-compression encoder, that constant cache is what enables single-pass multi-page transcription.

R-SWA is general-purpose parsing attention — beyond OCR it applies to ASR, translation, and similar long-horizon copy tasks.

Download or use

Reasoning for

Long-document OCR for the knowledge base. Where a chat-clipper or a generic OCR chokes on multi-page scans, a constant-KV-cache model parses an entire PDF in one shot — relevant to feeding scanned books/papers into the vault and to the long-document handling OpenKB / PageIndex solve from the retrieval side. Self-hostable (Transformers or SGLang), so document parsing stays local. For a full layout-aware document pipeline (tables → HTML, formulas → LaTeX, format conversion across PDF/DOCX/PPTX/XLSX) rather than a raw OCR model, see MinerU.

🧩 Inference

Two paths, both on NVIDIA GPUs (tested python 3.12 + CUDA 12.9):

  • Transformers — AutoModel.from_pretrained('baidu/Unlimited-OCR', trust_remote_code=True), then model.infer(...) for a single image (configs: gundam = base_size 1024 / image_size 640 / crop; base = 1024/1024/no-crop) or model.infer_multi(...) for multi-page / PDF (base only). PDFs are rasterized page-by-page with PyMuPDF first.
  • SGLang — launch an OpenAI-compatible server (sglang.launch_server --model baidu/Unlimited-OCR --enable-custom-logit-processor ...), stream requests with a DeepseekOCRNoRepeatNGramLogitProcessor. infer.py auto-starts the server and fans out concurrent requests over an image dir or PDF (--concurrency 8 --image_mode gundam).

Long-horizon parsing demo (animated): long-horizon-ocr.gif — not embedded (78 MB).

Alternatives considered

  • DeepSeek-OCR / DeepSeek-OCR-2 — the baseline Unlimited-OCR extends; growing KV cache on long output.
  • PaddleOCR — classic pipeline OCR.

Resources

Template: tool

Graph View

Backlinks