Granite-Docling-258M

Granite-Docling-258M

Converts document page images into structured text, with equations, layout and questions about page elements, in a 258M model from IBM Research. For people who build document conversion with Docling.

Made by ibm-granite on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Granite Docling is a small multimodal model from IBM Research that turns document page images into structured text. It has 258M parameters and fits into the Docling document pipelines. Its vision part is siglip2-base-patch16-512, and its language part is a 165M Granite model.

Who it suits: It suits developers who need to convert scanned or digital documents into formats such as Markdown and HTML. It also suits teams that already use Docling and want a model that plugs in.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Document-pipeline developers | Convert pages to Markdown or HTML | A converter for a folder of reports
Research | Data teams | Turn papers with equations into clean text | Maths-heavy pages with formulas intact
Back office | Operations teams | Read page structure from document images | Check which elements a page holds

If you only work with plain text files, you can skip it. It is for page images and documents.

Where to use it:

• The Docling library, the easiest route, which downloads the model by itself
• Plain transformers, vllm, onnx or mlx-vlm, which the card lists for direct inference
• Windows, macOS or Linux with Python
• A Mac with Apple's MPS accelerator, through a Docling option for it

Title: Jobs it does
Use case | Who | What they get back
Page conversion | Developers | A page in Markdown or HTML
Equation recognition | Researchers | Formulas detected and formatted more accurately
Region inference | Pipeline builders | Output for a chosen region of a page
Element questions | Analysts | Answers about which elements exist and in what order

How to set it up:

• Python on Windows, macOS or Linux
• The Docling library for the easiest route, installed through pip
• For plain transformers, the three prerequisite packages below
• About 515 MB of space for the BF16 weights

1. Install the Hugging Face Hub tools and download the model files.
2. Or install Docling with pip, which fetches the model for you.
3. For direct use, install the three packages the card lists.

pip install -U huggingface_hub
hf download ibm-granite/granite-docling-258M

pip install torch
pip install docling_core
pip install transformers

How to use it: With Docling, you run its command line tool and pick the vlm pipeline with the granite_docling model. You can write HTML, Markdown or split-page HTML, and ask for a layout overlay. The card's examples use a web address, so it is left out here.

Title: Docling options in the card
Tool | What it does
--to | Picks an output format, such as html or md
--pipeline vlm | Uses the vision-language pipeline
--vlm-model granite_docling | Selects this model
--show-layout | Draws the layout on the split-page output

hf download ibm-granite/granite-docling-258M

Example: You point Docling at a research paper with the granite_docling model. It reads each page image and writes Markdown plus an HTML version, with the layout shown if you ask. The card shows this on a paper in a picture.

For newcomers and experts: If you are new, this is a small AI that reads a photo or scan of a document page and writes it out as clean text. The first small thing to try is the Docling command on one short document.

If you know the field, it builds on the Idefics3 design, with the vision encoder swapped for siglip2-base-patch16-512 and a Granite 165M language model. It adds more accurate equations and tends to avoid endless output loops. The trade-off is language range: Japanese, Arabic and Chinese support is experimental.

Give Docling a page image or document
The Granite Docling model reads each page
Its output converts into Markdown or HTML
You check the result against the page

Good to know:

Title: Weight size in megabytes
BF16: 515 MB
4-bit (estimate): 129 MB

• Language: the main language is English, with Japanese, Arabic and Chinese in an experimental state.
• Modes: it can read a whole page at once or a region guided by a bounding box.
• Release: IBM Research released it in September 2025.
• Variants: other people have published 16 quantized and 16 fine-tuned versions.
• Demo: the card links a hosted demo on Hugging Face Spaces.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named, and the model is small enough for modest hardware.
• Data it reaches: it reads the documents you give it, and you download files from Hugging Face.
• Not for: pages that need languages beyond the ones the card lists.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 258M
• Weights: about 515 MB in BF16; about 129 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 245,919
• Likes: 1,295
• License: apache-2.0
• Languages: en
• Quantized versions: 16
• Fine-tunes: 16
• Tags: transformers, safetensors, idefics3, image-text-to-text, text-generation, documents, code, formula, chart, layout, table, document-parse

Links

• Model: https://huggingface.co/ibm-granite/granite-docling-258M
• Files: https://huggingface.co/ibm-granite/granite-docling-258M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ibm-granite/granite-docling-258M

Then follow the usage in its model card. Nothing in this kit runs by itself.