Baidu released Unlimited OCR 2, a 3.34-billion-parameter open-weight model under the MIT license for parsing very long documents. Its Sparse Reference Sliding Window Attention (SR-SWA) condenses text-visual attention into a fixed-size set of working references, pairing a local window with top-k block selection. Baidu reports 7.0x faster per-step decoding than full attention on 100-page documents and accurate parsing up to 300 pages on a single GPU, though the benchmark results are self-reported.
