Granite 3.2 Vision

tl;dr: a 2B vision-language model that fits in 2.4GB and reads documents like a structured-data extractor. Designed for “give me the table from this screenshot,” not for general multimodal chat.

What it is

granite3.2-vision is IBM’s smallest vision model in the Granite family. It is a vision-language fine-tune of a Granite text model: image and text go in, structured text comes out. The training corpus mixes public and synthetic instruction-following data tuned for document understanding — tables, charts, infographics, plots, and diagrams.

The default tag on the Ollama page is granite3.2-vision:2b, listed at 2.4GB on disk with a 16K context window and Text + Image input. The page lists vision and tools as capability tags. The library page does not state a license on this model.

Pull

ollama run granite3.2-vision
# or pin the tag:
ollama run granite3.2-vision:2b

What you get

• Text + image input. Pass images alongside text in the prompt.
• Tools capability listed on the Ollama page, so it slots into agent loops that need structured extraction.
• Granite LLM backbone — same family as the text-only granite3.3 and the larger Granite 4 line.
• Compact weights (2.4GB on disk). The whole thing fits on a laptop.

Hardware

Gotchas

• Requires Ollama 0.5.13 or newer — the library page calls this out explicitly. Older builds will fail to load the vision pipeline.
• 16K context, not 128K. This is a document-understanding model, not a long-context summarizer.
• Output is text only. No image generation.
• The Ollama library page does not state a license. IBM’s broader Granite models ship under Apache 2.0 upstream; check huggingface.co/ibm-granite/granite-vision-3.2-2b for the canonical terms before commercial deployment.

Related: Ollama for the runtime, llama.cpp for the engine, Granite 3.3 for the text-only counterpart, Gemma 4 for another open multimodal family on Ollama.

Untested here — facts from ollama.com/library/granite3.2-vision fetched 2026-09-10.