FULL STORY

Baidu Open-Sources Unlimited OCR 2

Baidu released and open-sourced Unlimited OCR 2, extending long-document parsing to thousands of pages. Follow-up benchmarks show 5.6x faster generation on 100-page documents with much lower VRAM usage.

2026-10-26 ~ 2026-10-27 · 2 episodes · 9 posts

Episode 1 · Baidu Open-Sources Unlimited OCR 2 for Thousand-Page Document Parsing (2026-10-26, 5 posts)

Baidu released and open-sourced ERNIE Unlimited OCR second generation on October 26-27, pushing the limit of ultra-long document OCR parsing from dozens of consecutive pages to the thousands-of-pages level. The company positions it as an unlimited-length document parsing solution for the "OCR-3.0 era."

Confirmed

  • The core mechanism is Unlimited-Horizon Decoding, paired with the newly proposed SR-SWA mechanism: text-vision attention interactions are compressed into a fixed-size set of working references, so attention computation for long documents no longer scales linearly with length and latency stays nearly constant.
  • Demonstrated capabilities: end-to-end one-pass parsing of the 1200+ page three-volume "The Three-Body Problem" and the 3600+ page full seven-book "Harry Potter" series.
  • Comparison data relayed by @xiaohu: KV cache dropped from 47GB to 1.77GB, a significant reduction.
  • The model has been released on Hugging Face as Unlimited-OCR-2, an image-text-to-text pipeline model with weights provided in safetensors; it can be loaded via transformers (flagged as customcode).

Why it matters

  • Ultra-long document parsing has long been a bottleneck for OCR and multimodal models: as documents grow, attention computation explodes, and previous approaches had to process page by page and stitch results together. With SR-SWA, Unlimited OCR 2 largely decouples compute and latency from document length, paving a viable path for one-shot structured parsing of entire books, lengthy contracts, and long reports.
  • Releasing the weights as open source means the community can deploy and build on the model directly, potentially upgrading applications such as long-document RAG and archive digitization.

Episode 2 · Unlimited OCR Gen 2 Runs 5.6x Faster on Long Documents with Far Less VRAM (2026-10-27, 4 posts)

Wenxin's Unlimited OCR Gen 2 processes 100-page documents 5.6x faster, improves multi-page scores from 61.89 to 84.59, and slashes cache VRAM from 47GB to 1.77GB by keeping only the active segment in memory, mimicking how humans transcribe text.