Unlimited-OCR: Parsing a 40-Page PDF in One Pass Without Your GPU Melting

Short summary

Baidu's Unlimited-OCR is a 3B-parameter MoE model (500M activated) that parses 40+ page PDFs in a single forward pass using Reference Sliding Window Attention, which keeps the KV cache constant-size by evicting old generated tokens while retaining permanent access to image tokens. It scores 93.23 on OmniDocBench v1.5, is MIT licensed, and runs on a 12GB+ NVIDIA GPU via Transformers, SGLang, or Docker. The R-SWA mechanism is general-purpose and applicable to any long-horizon transcription task including ASR.

• •Unlimited-OCR uses Reference Sliding Window Attention to keep KV cache constant, enabling single-pass multi-page OCR
• •3B param MoE (500M activated), MIT licensed, 93.23 on OmniDocBench — 6.2 points above DeepSeek-OCR baseline
• •Three deployment paths: Transformers script, SGLang server, Docker image; requires 12GB+ NVIDIA GPU

Generated with AI, which can make mistakes.