LLaVA

What LLaVA is for

LLaVA takes images as well as text. Microsoft + UW shipped it in 2023-12 with 3 sizes (7B, 13B, 34B), 4K of context and a Apache 2.0 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 open multimodal chat.

Sizes and memory

The 7B needs about 6 GB of weights at Q4_K_M and the 34B about 22 GB, a spread of 16 GB across the family. The catalogue's default pick is 13B. 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
7B	6 GB	9 GB	NVIDIA Jetson Orin Nano 8GB
13B	10 GB	15 GB	NVIDIA RTX 3060 12GB
34B	22 GB	33 GB	Apple M4 Max

Hardware that runs LLaVA

Every one of the 19 cards with dedicated memory in our hardware guide holds at least one size of LLaVA at Q4_K_M. The cheapest that does is the NVIDIA Jetson Orin Nano 8GB at about $199, running the 7B.

The largest size any of them holds is the 34B at 22 GB, on the Apple M4 Max (36 GB at 410 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 LLaVA: vision Q&A, image chat, doc understanding. It is tagged vision, multimodal, image, chat.

Where LLaVA sits in the LLaVA family

LLaVA is the only LLaVA entry in our catalogue, so there is no in-family alternative to weigh it against; the comparison to make is against the models listed at the foot of this page.

The Apache 2.0 licence puts no usage ceiling on it, which matters if the thing you are building has users. That is the practical difference between this and a vendor community licence.

Context window: 4K

4K tokens is a short window by 2026 standards — on the order of 3,000 words at the usual three-quarters-of-a-word-per-token ratio. That is one conversation turn with a document, not a session with a codebase. Anything longer has to be retrieved and injected rather than held.

How to run LLaVA 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 LLaVA 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 "LLaVA 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.