Baidu's open-source OCR model that parses very long multi-page documents in one pass with near-constant decoding cost.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Unlimited OCR 2 is an open-source document parsing model from Baidu, a follow-up to the original Unlimited-OCR released in June 2026, which was positioned as a step beyond DeepSeek-OCR. It takes page images and outputs parsed text for each page, and can process all pages of a document in a single request so that cross-page context is preserved.
Version 2 introduces Sparse Reference Sliding Window Attention (SR-SWA), which compresses text-visual attention over long sequences into a fixed-size set of working references (for example 512 tokens). According to the README, this keeps per-step decoding latency and memory use nearly constant regardless of document length. The project reports state-of-the-art results on MPDocBench and on MarathonDocBench, a new ultra-long document benchmark released by Baidu. YouTube reviewers describe tests on documents of around 5,000 pages and a model size of about 6.9 GB.
It can be run with Hugging Face Transformers (multi-page infer_multi and single-page infer modes) or served with a customized SGLang build that supports page-window scheduling. vLLM support, ms-swift training support, ModelScope and Baidu Cloud availability, and a Hugging Face Space demo are also listed. The documented tests used an NVIDIA A800 80GB GPU, so a CUDA GPU is expected.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source OCR model with no page limits or growing attention cost. It parses very long documents in one pass and held up on 5,000-page tests. The creator says he tested it and found it impressive.
Document reading model that processes multiple pages together to keep context across long documents. Beats several larger open models on parsing benchmarks; only 6.9GB.
Closest neighbours by meaning, not just by category.
Model · Baidu
Baidu's open-weights vision-language OCR model for parsing very long, multi-page documents.
Model · Baidu
Baidu's open-weights 280B-parameter sparse MoE base model (14B active) with 1M-token context, released pre-post-training.
Model · NVIDIA
NVIDIA's open multimodal model (120B total, 12B active parameters) for image, video and document understanding.
Model
Open-source append-only streaming speech recognition model with configurable low-latency decoding, English and Mandarin.
Model · AIDC-AI
Open multimodal embedding models from AIDC-AI that map text, images, video and documents into a shared vector space.
Model · Rednote
Open MoE language model from Rednote (Xiaohongshu) with 208B total and 16B active parameters.
