Radar ·
Baidu releases Unlimited OCR 2, open-source OCR model for long documents
Baidu announced Unlimited OCR 2, an open-source document parsing model that condenses text-visual attention into fixed-size working references, according to the company.
According to the briefing published on October 27, 2026, Baidu announces Unlimited OCR 2, an open-source document parsing model. The mechanism described, called SR-SWA, condenses text-visual attention interactions into fixed-size working references. According to the company, this keeps latency and memory nearly constant in OCR of long, multi-page documents. The source presents this result as the vendor's claim, without benchmark details in the original text.
For teams studying OCR of extensive batches, the material is relevant as a reference for techniques that limit attention and memory. When using AI to study or apply this content, do not paste real documents containing CPF, CNPJ, e-mail addresses or names into external tools; use fictitious examples. Consult the original announcement on Baidu's page and verify license, requirements and results before any adoption. Rota Nacional does not implement SR-SWA and does not claim to solve this engineering problem.
Original source: Baidu lança o Unlimited OCR 2 de código aberto para documentos longos