Unlimited OCR Works

The article introduces Unlimited OCR, a novel model that enhances OCR performance by utilizing a Reference Sliding Window Attention (R-SWA) mechanism to maintain a constant KV cache during decoding, thereby reducing memory consumption and improving efficiency for long sequences. This model, based on DeepSeek OCR, can transcribe up to 32K tokens in a single pass, making it suitable for extensive document processing. R-SWA's general applicability extends beyond OCR to other tasks like ASR and translation, providing a significant advancement for practitioners in the field.