Baidu releases Unlimited-OCR-2 with unlimited-length document parsing

Baidu published the Unlimited-OCR-2 model on HuggingFace for the image-text-to-text task. The release took place on October 26, 2026. The main change is the move from a long parsing horizon to an unlimited one thanks to the Sparse Reference Sliding Window Attention (SR-SWA) mechanism, which compresses the interaction of text and visual features on ultra-long sequences.

What it means

The model contains 3.3 billion parameters, with weights published in safetensors format. The largest file takes up 4.7 GB, and the entire repository is 6.2 GB. The license is MIT. It runs via transformers, and vLLM support is also claimed. There are no quantized builds on the platform yet. Running it locally will require a GPU with enough VRAM for 6.2 GB of weights at full precision plus memory for activations and the attention cache. The model can be run on a CPU, but processing long documents will be noticeably slower. You can follow the appearance of quantized versions and other news on this topic in the CPU3D news feed. The code is open source, so if you want, the model can be fine-tuned or adapted to your own document formats.