Granite Docling 258M

Granite-docling-258M is a compact image-plus-text-to-text vision-language model developed by IBM Research and released on September 17, 2025 under Apache 2.0. It is built on the IDEFICS3 architecture with a SigLIP2 vision encoder and a Granite 165M language model, and it is designed to live inside the Docling document-conversion toolkit rather than act as a general chatbot.

It consolidates what used to be several single-purpose models into one 258M-parameter VLM: full-page PDF conversion to Markdown/HTML, table-to-OTSL conversion, formula-to-LaTeX, code-block extraction, OCR with region selection, and even document-element question answering. It runs at about 0.35 seconds per page on an A100 and is also available in ONNX and MLX formats for edge deployment.