Knowledge Base

Baidu Unlimited-OCR: Open-Source One-Shot Long-Document Parsing

2026-07-19inferencellmocr

Baidu Unlimited-OCR: Open-Source One-Shot Long-Document Parsing

Source: https://x.com/i/status/2078774556249186345

📌 Unlimited-OCR is Baidu’s MIT-licensed ~3B vision-language OCR model that parses multi-page documents in a single forward pass via Reference Sliding Window Attention. It scores ~93.23% on OmniDocBench v1.5 and runs fully offline on common local stacks as a free alternative to cloud OCR.

📄 One-Shot Long Docs

Reads multi-page PDFs and image stacks in one pass under a 32,768-token context. R-SWA keeps the KV cache roughly constant as documents grow, avoiding page-by-page context loss and memory blow-up.

🧠 Built on DeepSeek-OCR

Keeps the encoder’s high image compression while replacing the decoder’s attention with Reference Sliding Window Attention. Reports only ~500M parameters active at inference despite ~3B total size.

💻 Local & Offline First

Runs on Transformers, vLLM, SGLang, Docker, Ollama, and llama.cpp. Community GGUF/NVFP4 quants and ~7–8.5 GB BF16 VRAM make single-GPU local use practical for privacy-sensitive or air-gapped workflows.

📊 Strong Benchmarks

About 93.23% overall on OmniDocBench v1.5—roughly six points above DeepSeek-OCR—with low error reported even past dozens of pages. Aimed at scanned/image PDFs and structured markdown-style output.

🆓 Free Forever (MIT)

MIT license and open weights undercut paid cloud OCR (often ~$1.50–$15 per 1,000 pages). Hugging Face shows 2M+ monthly downloads; GitHub and HF Spaces demos accelerate adoption.

Key facts

Fact Value
Model size ~3B total params (~500M activated at inference)
Context / method 32K tokens; one-shot parsing via R-SWA (constant KV cache)
Benchmark ~93.23% OmniDocBench v1.5 (~+6 pts vs DeepSeek-OCR)
License / release MIT; published ~June 22, 2026 (arXiv:2606.23050)
VRAM (BF16) ~7–8.5 GB; lower when quantized (GGUF/NVFP4)
Adoption ~2M+ monthly HF downloads; local runtimes + demos

Details

Unlimited-OCR is Baidu’s open-source vision-language OCR model designed for long, multi-page documents. Unlike pipelines that slice pages (losing cross-page context) or send jobs to paid cloud APIs, it processes image stacks and scanned PDFs in a single forward pass while keeping memory growth under control through Reference Sliding Window Attention.

Architecturally it extends DeepSeek-OCR: the encoder retains aggressive image compression, and the decoder’s attention is swapped for R-SWA so the KV cache stays roughly constant as page count rises. That design fits a standard 32K-token context and reduces the quality drift common in page-by-page OCR. Total size is about 3B parameters, with reports of only ~500M active at inference—compact enough for single-GPU local use, especially once community quantizations are applied.

On OmniDocBench v1.5 it scores about 93.23%, roughly six points above the DeepSeek-OCR baseline. Released under MIT around June 22, 2026 (arXiv:2606.23050), it targets privacy-sensitive, high-volume, and air-gapped workflows that need structured markdown-style output without per-page cloud fees. A July 2026 viral X post highlighting its small footprint and free local use helped drive renewed interest and multi-million monthly downloads on Hugging Face.

Sources