Baidu published Unlimited OCR 2 this week, a 3B-parameter document model under an MIT license that it says parses documents of hundreds of pages in a single request.
The obstacle it targets is the one that breaks ordinary OCR at scale. Standard vision-language attention grows quadratically with sequence length, so memory and per-token latency climb as a document gets longer until a scanner either slows to a crawl or runs out of room.
Baidu’s answer is Sparse Reference Sliding Window Attention, or SR-SWA. Instead of attending across the whole sequence, it compresses text-visual interactions down to a fixed working set of roughly 512 tokens, which bounds the computation and holds step time nearly constant regardless of document length.
The company calls the result unlimited-horizon parsing within a finite context window, and says it sets state of the art on both short benchmarks such as MPDocBench and extreme ones aimed at hundred-page inputs.
Alongside the weights, Baidu released MarathonDocBench, a new benchmark built for ultra-long documents, where reports put the model near 77.1.
Weights are available on Hugging Face and ModelScope, with inference supported through vLLM and SGLang and a hosted demo running in a Hugging Face Space. The model descends from DeepSeek-OCR, which Baidu credits in its documentation.
The practical appeal is unglamorous and large. Filings, scanned books and long regulatory reports are exactly where current document pipelines break, and a model that keeps its cost curve flat changes what can be processed in one call.
