DeepSeekOCR cuts token counts for long-context LLMs

DeepSeek, a Hangzhou-based artificial intelligence startup, on Monday released DeepSeekOCR, an open source multimodal model that converts large documents into compressed visual representations to reduce the number of tokens required for processing. The company said the approach can cut token counts by seven to 20 times, enabling large language models to handle much longer contexts without a proportional rise in computing cost, and the model is available on developer platforms including GitHub and Hugging Face.

DeepSeekOCR combines a vision encoder called DeepEncoder with a decoder named DeepSeek3BMoEA570M, a mixture-of-experts model with about 570 million parameters. The encoder transforms text and structured content into high-resolution images or 2D grids that the decoder reconstructs into text. According to the company, the system preserves complex elements such as tables, charts, formulas and diagrams while reducing the token load required for downstream LLM processing. The startup reported benchmarks showing the model outperformed several leading optical character recognition systems on OmniDocBench while using far fewer tokens.

In tests cited by the developer, when visual...