According to Bijie.com, the DeepSeek-AI team has released the paper "DeepSeek-OCR: Contexts Optical Compression," proposing a new method for compressing the context of long text using visual modalities. The Hugging Face page indicates that the model has 3B parameters.