IBM today announced Granite-Docling-258M, an ultra-compact open-source vision-language model designed to convert documents into machine-readable formats while preserving layout, tables, equations, lists and other structural elements. the model has 258 million parameters and is available on Hugging Face under a standard Apache 2.0 license. IBM positions Granite-Docling as a product-ready successor to the experimental SmolDocling-256M-preview model and highlights cost efficiency and one-shot parsing as core benefits for document conversion and downstream retrieval augmented generation workflows.
Granite-Docling adopts a Granite 3-based language backbone and the SigLIP2 visual encoder while maintaining the general SmolDocling methodology. the model targets common failure modes observed in the preview release by applying extensive dataset filtering and cleaning to remove inconsistent or ambiguous annotations and to mitigate instabilities such as token repetition loops. central to Granite-Docling’s approach is DocTags, a universal markup format developed by IBM Research that separates textual content from document structure and encodes element locations and relationships. this format is optimized for language model readability and can be converted to Markdown, JSON or HTML or fed into Docling pipelines. IBM also documents experimental multilingual capabilities, listing Arabic, Chinese and Japanese as initial targets while noting that multilingual support is early and not yet validated for enterprise readiness. performance evaluations are provided on Granite-Docling’s Hugging Face model card.
Granite-Docling is intended to complement rather than replace the Docling library. IBM recommends using the model within Docling for ensemble pipelines, integration with external services and agentic workflows. planned initiatives include continued development of the open-source docling-eval package and curated evaluation datasets, larger Granite-Docling model sizes of about 512 million and 900 million parameters (kept below 1 billion), and DocTags compatibility with IBM watsonx.ai models and tokenizers. IBM provides links to the Hugging Face model card, docling.ai, tutorials and workshops for getting started with Granite-Docling and the Docling ecosystem.
