IBM released Granite-Docling-258M as a compact document-conversion VLM
IBM released Granite-Docling-258M on September 24, 2025, an Apache-2.0 258M vision-language model for structured document conversion.
Historical reporting: IBM announced Granite-Docling-258M on September 24, 2025 as an Apache-2.0 open vision-language model for document conversion. IBM’s release and its publisher model card describe its 258M size, document-focused task, stated handling of layout, tables, equations, lists and code, and available local-serving paths. IBM’s efficiency and comparative statements are supplier claims. They do not establish output accuracy, hallucination rate, layout fidelity, throughput, privacy handling, model suitability or production support for another corpus or serving configuration.
Key takeaways
- IBM released a 258M document-focused vision-language model that aimed to preserve document structure alongside text.
- Granite-Docling was positioned as a single-model path beside Docling’s configurable conversion pipelines.
- Its compact size and Apache-2.0 availability do not remove the need to test layout, table, equation and provenance fidelity.
IBM announced a 258M document-conversion model rather than a general OCR endpoint[source 1][source 2]
IBM announced Granite-Docling-258M on September 24, 2025 as an open-source vision-language model for converting documents into machine-readable formats. The announcement says the 258M-parameter model was available through Hugging Face under Apache 2.0 and describes its target as preserving layout, tables, equations, lists, inline and floating math, and code. IBM positioned the release beyond plain text extraction: the intended output should retain structural information useful to later processing. That makes the event a document-understanding model release, not proof that every output retains the meaning of its source page. A footnote may attach to the wrong paragraph, a table header may be flattened or a formula may lose a symbol. A workflow using converted material for retrieval or automation still needs ways to inspect the original page and resolve ambiguous output.
IBM described Granite-Docling as a one-shot model path with a stated architectural lineage[source 1]
IBM said Granite-Docling was developed by the team behind the Docling library, which supplies document-conversion tools, models and a command-line interface. The announcement distinguishes configurable Docling ensemble pipelines from Granite-Docling’s single 258M-parameter VLM that parses and processes a document in one shot. It also says the new model evolved from the experimental SmolDocling-256M-preview: IBM replaced the earlier SmolLM-2 language backbone with a Granite 3-based architecture and replaced the SigLIP visual encoder with SigLIP2 while retaining the general methodology. This is useful lineage context, not evidence that one model universally supersedes a multi-stage pipeline. A single-model route can simplify a stack while moving more responsibility into model inference. Operators still need to choose page rendering, prompts, output schema, validation, fallbacks and review conditions.
The publisher record identified an image-to-text artifact and multiple local-serving routes[source 2]
The IBM Granite publisher page identifies granite-docling-258M as an image-text-to-text Transformers and Safetensors model with document, formula, chart, OCR, layout, table and extraction tags. Its model card gives loading examples for Transformers and documents vLLM and SGLang serving routes. These are meaningful integration facts, but they are not a performance certificate. A model can be loaded through several runtimes while tokenizer versions, image resolution, batching, maximum output tokens, prompt template, device memory and response parsing differ. A team that converts a PDF should record which page renderer and model revision were used, then preserve source-page references in the output. That evidence makes it possible to distinguish a source document from the model’s serialization of it and to investigate whether a later answer relied on a missed or invented structural element.
A compact model changes conversion economics without making a converted document self-authenticating[source 1][source 2]
IBM described Granite-Docling as cost-effective and said its capabilities rival systems several times its size. The source provides a useful explanation for why a smaller document-specific model could be attractive: document conversion can run before retrieval, indexing or analysis, so latency and compute compound across a large corpus. Yet the same placement amplifies extraction errors. If a model wrongly reconstructs a table, downstream search can return a confident answer with no obvious trace of the original page. Review should therefore focus on the features IBM named: text, layout, tables, equations, code and lists. Use representative documents, measure field-level correctness and require a page or region reference for high-impact extracted claims. IBM’s September release documents a model and its stated scope; it does not make generated Markdown a substitute for the original record.
Sources
Continue reading
Related reading
News analysisAi2 released olmOCR as an open 7B vision model for PDF extraction3 min read
News analysisNASA and IBM launched a lunar foundation model with code and model access for science work3 min read
News analysisIBM and NASA opened Surya for heliophysics research, with a NASA technical record of its solar inputs3 min read