Baidu's Unlimited OCR 2 reads the 1,200-page Three-Body trilogy in one pass, scoring 77.1 vs 33.7
heyshrutimishra · x · 2026-10-27
Baidu's new Unlimited OCR 2 claims to transcribe the entire 1,200+ page Three-Body trilogy in a single pass. The hard part isn't per-page OCR but cross-page consistency over long documents: MiniMax-M3 fails past 223 pages and Gemini 3.1 Pro / Qwen3.8-Flash-Next degrade with length, while Unlimited OCR 2 scores 77.1 on MarathonDocBench vs 33.7 for the runner-up.
The trick is Sparse Reference Sliding Window Attention (SR-SWA) — eyes on the current page, glancing back at relevant earlier pages — so per-step decode cost barely grows with length (7× faster than full attention at 100 pages). Its predecessor has 10M+ Hugging Face downloads and 28K GitHub stars. Caveats: numbers are from Baidu's own report on clean printed text; independent tests on messy scans, handwriting, and mixed-language PDFs are still needed.
Related event: Baidu Open-Sources Unlimited OCR 2 for Thousands-of-Pages Document Parsing(9 posts)→
More from Models
- Rumor: OpenAI's 25-page proof of Sogge smoothing conjecture bypasses 20 years of work — teortaxesTex · 2026-10-27
- Cohere releases Qwen traces: anomaly is real, but backdoor intent claims have no evidence — basedjensen · 2026-10-27
- Fields medalist Yu Deng calls out OpenAI's double standard on distillation — teortaxesTex · 2026-10-27
- Commentary: OpenAI won't release reasoning traces because the reasoning process is the real product — max_paperclips · 2026-10-27
- 229k-sample replication of Cohere's Qwen 3.8 malicious PUT claim raises more questions — builderjaydub · 2026-10-27
- Chinese LLMs may never shed "I'm Claude" contamination without abandoning distillation — teortaxesTex · 2026-10-27