Baidu publishes MIT-licensed Unlimited-OCR-2 weights, code and inference instructions; performance claims await independent reproduction.
Baidu releases Unlimited-OCR-2 weights for ultra-long documents
2026-10-26Verified announcementTRIAL
Baidu published MIT-licensed Unlimited-OCR-2 weights, code, a demo and a new ultra-long-document benchmark. Its Sparse Reference Sliding Window Attention keeps a fixed set of working references while parsing longer documents, and the model supports Transformers plus a customized SGLang path. Baidu reports leading results on multi-page benchmarks, but the comparisons have not been independently reproduced.
Availability: MIT-licensed weights, code, demo, ModelScope and Baidu Cloud access
What this means for you
For document-RAG teams, a bounded-memory OCR path could reduce the need to split long reports before parsing and preserve cross-page context. The practical question is whether accuracy, latency and memory hold on real scans outside the vendor benchmark.
Evidence & related material
Technical details & scores
| Method | Sparse Reference Sliding Window Attention (SR-SWA) with a fixed working-reference set such as 512 tokens |
|---|---|
| Availability | MIT-licensed weights, code, demo, ModelScope and Baidu Cloud access |
| Runtimes | Transformers and a customized SGLang serving path |
| Documented Test Hardware | NVIDIA A800 80GB |
| Adoption Snapshot | 513 downloads and 189 likes when checked |
| Evidence Limit | Benchmark leadership and scaling claims are vendor-reported; no independent reproduction found |
| Time Basis | Hugging Face repository created 2026-10-26T03:57:46Z with model uploads that day; the release list labels the public launch 2026-10-27 without a time, so this is a recall-audit late discovery |
Technical impact 4
Engineering value 5
Adoption signal 2
Maturity 3
Verification cost 3
Risk 3