DeepSeek has released DeepSeek-OCR, an open-source multimodal model that compresses text by rendering it as images, on GitHub and Hugging Face. The model pairs a vision encoder with a 3B-parameter mixture-of-experts decoder (570M active) and, according to the accompanying paper, achieves 97% OCR precision when the ratio of text tokens to vision tokens is under 10x. At a 20x compression ratio, accuracy falls to about 60%. DeepSeek frames the work as an initial investigation into “contexts optical compression”—using 2D visual representations to shrink long contexts for large language models. The company says this could lead to significant token reduction, potentially lowering memory and compute costs for long-context tasks. On benchmarks, DeepSeek-OCR outperforms GOT-OCR2.0 with 100 vision tokens per page and surpasses MinerU2.0 using fewer than 800 vision tokens, versus MinerU2.0’s 6000+. In production, it can generate training data for LLMs and VLMs at over 200,000 pages per day on a single A100-40G GPU. The release continues DeepSeek’s push for efficiency and cost reduction in AI models.