mPLUG-DocOwl is an open-source multimodal large language model family from Alibaba’s X-PLUG team, built specifically for understanding documents. Released under the Apache-2.0 license with more than 2,400 GitHub stars and updates through 2025, it tackles the documents that general vision-language models struggle with: dense pages, tables, charts and forms.
General multimodal models downsample page images so aggressively that small text becomes unreadable. mPLUG-DocOwl instead uses a modularized design: a vision encoder that preserves document structure, an abstraction module that compresses visual tokens while keeping layout information, and a language model that reasons over the result. Successive versions, from the original DocOwl through DocOwl 1.5 and DocOwl2, pushed accuracy on document VQA, chart understanding and table recognition benchmarks while cutting the visual token count, which makes inference faster and cheaper.
mPLUG-DocOwl is free and open source under the Apache-2.0 license, including the released model weights, subject to the base models’ terms.
The project suits researchers and engineers building document intelligence pipelines โ invoice reading, chart question answering, form extraction โ who want an open model they can fine-tune and self-host instead of relying on a proprietary API.
The DocOwl line has evolved through several generations โ DocOwl 1.5 unified single-image, multi-page and chart understanding in one model, and DocOwl2 pushed further on speed by compressing each page into a small set of visual tokens. The repository publishes training and evaluation code, instruction-tuning recipes and links to checkpoints, and reports results across a broad benchmark suite including DocVQA, InfoVQA, ChartQA and table recognition datasets, which makes the numbers easy to reproduce and compare against.