Baidu's Unlimited OCR 2 is a 3B vision-language model capable of parsing long documents efficiently.
Claims
Baidu's Unlimited OCR 2 is a 3B vision-language model capable of parsing long documents efficiently.
Parent: AIEntity: Unlimited OCR 2Impact: positiveDate: Oct 27, 2026 - Oct 28, 2026Target: Efficiency in document parsing
The Sparse Reference Sliding Window Attention mechanism in Unlimited OCR 2 keeps decoding costs almost flat as document length increases.
Parent: AIEntity: Unlimited OCR 2Sub-entity: Sparse Reference Sliding Window AttentionImpact: positiveDate: Oct 27, 2026 - Oct 28, 2026Target: Efficiency in decoding costs
Baidu's Unlimited OCR 2 has been tested and can parse all seven Harry Potter books, totaling 3,600 pages, in a single run.
Parent: AIEntity: Unlimited OCR 2Impact: positiveDate: Oct 27, 2026 - Oct 28, 2026Target: Document parsing capability
Source posts
Baidu has open-sourced Unlimited OCR 2, a 3B vision-language model that parses hundred-page documents in a single request. Its SR-SWA sparse attention keeps decoding cost almost flat as page counts grow. Weights are MIT licensed and it runs on a single GPU. https://www.marktechpost.com/2026/10/26/baidu-unlimited-ocr-2-open-source-long-document-ocr/ #AI #GenAI #AIresearch
0 boosts · 0 favs · 0 replies · Oct 27, 2026
#ai#genai#airesearch
Baidu has open-sourced Unlimited OCR 2, a document parser that reads documents of any length in one pass. It uses Sparse Reference Sliding Window Attention, keeping memory and latency nearly constant. Baidu says it parsed all seven Harry Potter books, 3,600 pages, in one run, and scores 84.6 on MPDocBench. Weights are MIT-licensed. https://pandaily.com/baidu-open-sources-unlimited-ocr-2-sr-swa-any-length-document-parsing-marathondocbench #China #Tech #AI
0 boosts · 0 favs · 0 replies · Oct 28, 2026
#china#tech#ai