The public model repository is MIT-licensed, ungated, and identifies the model as an image-to-text system using Transformers custom code. Its safetensors metadata lists 3.34 billion BF16 parameters and roughly 6.69 GB of model files.
The release introduces Sparse Reference Sliding Window Attention (SR-SWA). Baidu says it compresses text-visual attention for ultra-long sequences into a fixed-size set of working references, with 512 tokens given as an example. The model card says this keeps per-step decoding latency and memory footprint nearly constant as document length grows, and supports multi-page input in one request so it can use cross-page context.
The repository documents local use through Transformers and recommends SGLang for serving. It also includes vLLM-oriented instructions. Baidu claims state-of-the-art results on MPDocBench and MarathonDocBench, including hundred-page document benchmarks.
The source does not independently verify those benchmark claims, establish accuracy on particular scans or languages, or provide a comparison of operating cost and hardware requirements across document sizes.