Tag
DeepSeek-OCR is a 3B vision model using context optical compression for efficient document processing. Fine-tuning it on Persian text using Unsloth achieved an 88.26% improvement in character error rate, all open-source and runnable on a single GPU.