Unlimited-OCR 2 on a 16 GB GPU: it reads as well as Chandra 2, but it is not unlimited

On October 27, Baidu unveiled Unlimited-OCR 2, a 3.3 billion parameter model released under the MIT license. Its claim was that its sparse attention mechanism maintained nearly constant per-step decoding latency and memory footprint, regardless of document length. To test this claim, a comparison was made between Unlimited-OCR 2, Chandra 2, Tesseract, Marker 2.0, and DeepSeek-OCR, using PDFs of real Central Bank of Chile reports.

The median character error rate per page for Unlimited-OCR 2 was 0.12%, which was on par with Chandra 2 and superior to Tesseract, Marker, and DeepSeek-OCR. However, when running on a 16 GB GPU, Unlimited-OCR 2 struggled as the document size increased. With more pages in one request, memory usage rose, and time per token increased.

The transformers path, which kept every token in a dynamic cache, was faster than the transformers, but the 10-page window mode had issues. For large files, Marker proved to be the practical choice, completing a 74-page report in 12.6 minutes using 4.2 GB of the GPU. Chandra 2 took 46 minutes and 11 GB of GPU memory. Unlimited-OCR 2 is MIT-licensed, while Chandra 2's weights are only available under specific funding conditions.

The post also covered installation difficulties on NixOS and potential threats to validity, with full measurements, pre-registered sheet, and scripts available at https://efraingaray.com/en/blog/unlimited-ocr-2-16gb/.