Baidu releases Unlimited-OCR, building on DeepSeek-OCR

Category: ML research
Status: developing
Sources: GitHub, weights, discussion
Summary: Baidu published Unlimited-OCR on 2026-06-22, an MIT-licensed document-parsing model with weights on Hugging Face and ModelScope, framed around "one-shot long-horizon parsing" of multi-page documents. The README states the work builds on and aims to push DeepSeek-OCR "one step further," and the inference code handles single images and multi-page PDFs. The repository does not yet document model size, architecture, or benchmark numbers.
Why it matters: It lands the same week as Mistral OCR 4 and continues the move toward dense, structure-aware document OCR; an open-weight MIT release lets teams self-host the pipeline rather than meter a closed API.
Follow-up: Confirm parameter count, architecture, and benchmarks against DeepSeek-OCR once a model card or paper is published.