Llama 3.2 Vision

What is Llama 3.2 Vision?

Llama 3.2 Vision is Meta's open-weights language model with vision support. Released in 2024-09, it ships in 2 sizes (11B, 90B) and is licensed Llama 3.2 Community License. The most popular use case is vision + chat on desktop GPUs.

VRAM and hardware

The smallest 11B size needs at least 7 GB VRAM; the largest 90B needs around 7 GB.

Size	Min VRAM (Q4_K_M)	Recommended VRAM	Tokens/sec on 3090
11B	7 GB	11 GB	~23 tok/s
90B	55 GB	83 GB	~14 tok/s

How to run Llama 3.2 Vision locally

Option 1: Ollama (simplest)

ollama pull llama-3-2-vision
ollama run llama-3-2-vision

Option 2: Mullama (production)

mullama pull llama-3-2-vision
mullama run llama-3-2-vision

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "llama-3-2-vision gguf")
./llama-cli -m llama-3-2-vision.Q4_K_M.gguf -p "Hello, AI!"

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

from mullama import Model, Context
model = Model.load("llama-3-2-vision.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with Llama 3.2 Vision

• vision
• image Q&A
• doc understanding

Hardware it fits on

• Apple Silicon (8GB): 11B in MLX
• Apple Silicon (16GB): 90B in MLX
• Apple Silicon (32GB): 90B in MLX or GGUF
• Apple Silicon (64GB+): 90B in MLX
• RTX 3090 (24GB): 90B at Q4_K_M
• RTX 4090 (24GB): same as 3090, ~30% faster
• RTX 5090 (32GB): 90B at Q8_0
• 2× RTX 6000 Ada (96GB total): 90B at Q4_K_M

Related models in the Llama family

• Llama 3.1 — 8B, 70B, 405B · general-purpose chat, agentic workflows, code
• Llama 3.2 — 1B, 3B · low-end hardware, mobile, edge deployment
• Llama 3.3 — 70B · best-in-class 70B for desktops
• Llama 4 — Scout 17B (109B MoE), Maverick 17B (400B MoE) · vision + long context (10M tokens)

Other models you might consider

• Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware
• Qwen 2.5-Coder (Alibaba) — 0.5B, best open-weight code model in 2024-2025
• Qwen 3 (Alibaba) — 0.6B, best all-round local LLM in 2025-2026
• Qwen 3-Coder (Alibaba) — 30B, long-context code agent workflows
• DeepSeek R1 (DeepSeek) — 1.5B, reasoning, math, step-by-step problem solving
• DeepSeek V3 (DeepSeek) — 671B (37B active MoE), frontier-class open-weights LLM (needs data-center GPUs)