IBM Granite Docling 258M

Granite Docling 258M does one job: it looks at an image of a document page and writes down what is on it — not as prose, but as DocTags, a markup that records where each element sits and what kind of element it is. A table comes back as a table with its cells intact, an equation as an equation, a code listing as code with its indentation. That output feeds straight into the open Docling toolchain, which turns it into Markdown, HTML or JSON. This is the unglamorous half of document AI and it is used far more heavily than the headline models. The repository was pulled about 239,000 times in the thirty days to 1 September 2026, more than twice as often as Granite Vision 4.1 4B, IBM's current general-purpose image model, and more than Granite Guardian. Anyone building retrieval over PDFs has to solve page conversion before anything else works, and a 258-million-parameter model runs on a laptop. The architecture is a small assembly of borrowed parts: the Idefics3 layout, with the vision encoder swapped for SigLIP2 base at 512 pixels and the language half replaced by a 165-million-parameter Granite model. It is the production successor to SmolDocling-256M-preview, and IBM's own tables show where the gains landed. Table structure recognition on FinTabNet climbs from 0.82 to 0.97 and table content from 0.76 to 0.96 — the difference between a table you have to retype and one you can use. Code recognition improves by nearly an order of magnitude: edit distance drops from 0.114 to 0.013. Equation recognition, layout detection and full-page OCR all improve more modestly, and OCRBench rises from 338 to 500. The model also accepts region-restricted prompts: instead of converting a whole page, it can be pointed at one bounding box, and it can answer questions about a document's structure, such as which elements appear and in what order. IBM lists better resistance to the repetition loops that plagued the preview version. Limits worth knowing: the model is English, with Japanese, Arabic and Chinese marked experimental in the model card — no claim is made for Polish. The licence is plain Apache 2.0, with no territorial carve-out and no acceptable-use annex, so the weights can be used commercially and offline, which matters in exactly the setting this model belongs to: contracts, invoices and medical records that must never leave the building.

▸Videos

No videos yet.

▸Reviews