Unlimited OCR vs DeepSeek OCR: benchmarks and handwriting

Baidu's Unlimited OCR is the subject of this comparison with DeepSeek OCR. The speaker describes a 3 billion parameter model that replaces the baseline decoder's attention layers with reference sliding window attention. According to the presentation, this keeps the KV cache constant during decoding and lets the model process hundreds of pages in one pass with lower memory use and stable speed.

The reported OmniDocBench results are 93% for Unlimited OCR versus 87% for DeepSeek OCR on version 1.5, and almost 94% versus 90% on version 1.6. These are benchmark claims presented in the video, rather than an independent evaluation. The speaker says consumer GPUs should be able to run the model, but gives no specific GPU or memory requirement. Hugging Face Transformers integration and a Hugging Face Spaces demo provide routes to trying it; Google Colab is also mentioned as an execution option.

Two handwriting examples give a more practical view of the output. The speaker reports an accurate transcription of a handwritten quotation and discusses text extraction with bounding boxes. A less legible list produces several mistakes, including reading "Some friends" as "home friends." The speaker recommends validation for difficult handwriting, particularly medical transcription. For readers considering local AI for document processing, the review covers reported efficiency gains and a concrete reason to check extracted text.