Baidu Open-Sources Unlimited OCR 2: 270-Page 1899 Drug Manual Parsed in One Request
MaziyarPanahi · x · 2026-10-29
Baidu has open-sourced Unlimited OCR 2, a document parsing model built for virtually unlimited document lengths. It combines Sparse Reference Sliding Window Attention (SR-SWA) with Unlimited-Horizon Decoding, focusing compute on the most relevant visual references while releasing KV caches no longer needed.
Key numbers: near-constant per-step decoding latency and memory as documents grow; state-of-the-art on MPDocBench and MarathonDocBench; on 100-page documents, 7.0× faster per-step decoding than Full Attention and 5.6× faster than R-SWA, maintaining strong accuracy up to 300 pages.
A user tested it on a doctor's drug manual from 1899—all 270 pages in one request on a single GPU, preserving the original layout, including the cocaine entry listing morphine as its antidote.
More from Multimodal
- Higgsfield launches Selected showcase; first pick MADSEASON is fully open source — CurieuxExplorer · 2026-10-29
- Wistia launches Mochine, an AI video editor that adds motion graphics to talking-head videos — Scobleizer · 2026-10-29
- Fizgig H3 still-image latent node drops into any ComfyUI workflow — and runs faster — robertwellesley · 2026-10-29
- Guess-the-model Reddit post shows local AI photorealism that's hard to tell from real photos — Licovoda · 2026-10-29
- OpenAI's Terra called a Pareto inversion for visual perception, reshaping physical AI economics — arnabdotorg · 2026-10-29
- AssemblyAI-TTS-1 tops TTS speed at 162 chars/sec as Artificial Analysis refreshes leaderboards — ArtificialAnlys · 2026-10-29