Hands-on with Baidu's Unlimited OCR 2: whole docs in one pass, no page limit
Baidu_Inc · x · 2026-10-28
Fahdmirza tested Baidu's newly released Unlimited OCR 2 locally:
- Parses entire documents in a single pass with no page limit; a new SR-SWA trick focuses only on tokens that matter, keeping speed and memory nearly flat on long docs.
- Authors report big wins on hundred-page benchmarks; his tests parsed handwritten physics equations, math solutions, and invoice tables in seconds.
- Limits: mixed-language handwriting, ornate old print, and empty table rows can trip it up.
More from Multimodal
- ByteDance's Seedance 2.5 demo claims the gap between AI video and cinema is closed — CurieuxExplorer · 2026-10-28
- That checkers robot clip was Genie 4 rendering, not a real Apollo learning — AIFlow_ML · 2026-10-28
- Genie 4 can imagine a robot's whole day — but robots can't train in it yet — AIFlow_ML · 2026-10-28
- Agentic video editing goes pro: senior editors 5x output, costs to collapse — dotey · 2026-10-28
- Alibaba's OvisOCR3: a 0.8B model tops OmniDocBench at 96.74 via multi-teacher distillation + RL — ATH-MaaS · 2026-10-28
- CodeCap-Bench: executable code beats natural language as video captions — Jiafu Tang · 2026-10-28