Llama 3.2 Vision

What Llama 3.2 Vision is for

Llama 3.2 Vision takes images as well as text. Meta shipped it in 2024-09 with 2 sizes (11B, 90B), 128K of context and a Llama 3.2 Community License licence. Running a vision-language model locally costs more memory than a text model of the same parameter count, because the image encoder and the projected image tokens both occupy the context. The catalogue records its strength as vision + chat on desktop GPUs.

Sizes and memory

The 11B needs about 7 GB of weights at Q4_K_M and the 90B about 55 GB, a spread of 48 GB across the family. The catalogue's default pick is 11B. Recommended VRAM adds roughly half again on top of the weights for the KV cache and the runtime.

Size	Weights at Q4_K_M	Recommended VRAM	Smallest card in our guide that holds it
11B	7 GB	11 GB	NVIDIA Jetson Orin Nano 8GB
90B	55 GB	83 GB	Apple M3 Ultra

Hardware that runs Llama 3.2 Vision

Every one of the 19 cards with dedicated memory in our hardware guide holds at least one size of Llama 3.2 Vision at Q4_K_M. The cheapest that does is the NVIDIA Jetson Orin Nano 8GB at about $199, running the 11B.

The largest size any of them holds is the 90B at 55 GB, on the Apple M3 Ultra (192 GB at 800 GB/s). Bandwidth, not capacity, sets the speed once a size fits, so a card further down the list with faster memory will generate more tokens per second on the same weights than a larger, slower one.

What people use it for

The catalogue lists 3 recorded uses for Llama 3.2 Vision: vision, image Q&A, doc understanding. It is tagged vision, multimodal, image, tools.

Where Llama 3.2 Vision sits in the Llama family

Our catalogue holds 7 Llama entries. Ordered by release date, Llama 3.2 Vision (2024-09) is the 5th of them, and Llama 4 (2025-04) is newer.

The Llama 3.2 Community License is a vendor community licence rather than a standard open-source one: usable, redistributable with attribution, and carrying conditions that are worth reading before you build a product on it.

Context window: 128K

128K tokens is roughly 96,000 words at the usual ratio, which is a small repository or a whole book. Two cautions come with a window this size: the KV cache at full length can rival the weights for memory, and models rarely use the far end of their advertised context as well as they use the near end. Treat it as headroom, not as a promise.

How to run Llama 3.2 Vision locally

Our catalogue does not record registry tags, so look the current tag up in the Ollama library or on Hugging Face before pasting these. The commands below are the shape of the workflow, not a copy-and-paste recipe.

Option 1: Ollama (simplest)

# find the tag for Llama 3.2 Vision at ollama.com/library
ollama run <tag>

Option 2: Mullama (production)

mullama pull <tag>
mullama run <tag>

Option 3: llama.cpp (CLI)

# download a GGUF from Hugging Face, searching for "Llama 3.2 Vision GGUF"
./llama-cli -m model.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("model.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

Sources

Model record from src/data/models.json (2026-06-29); card capacities from src/data/gpus.json (2026-06-29). Upstream: Curated from ollama.com/library + community benchmarks + paperswithcode + huggingface.