The vision models use cross-attention layers added to the language backbone to process image features - a cleaner integration than adapter-based approaches.
# 11B vision model - runs on RTX 4090 (24GB) or M2/M3 Max Mac
ollama pull llama3.2-vision:11b
# 90B vision model - requires 2-4x A100s
ollama pull llama3.2-vision:90b
# Text-only on-device variants
ollama pull llama3.2:1b
ollama pull llama3.2:3b
Using Vision Capabilities
import ollama
# Image analysis
response = ollama.chat(
model="llama3.2-vision:11b",
messages=[
{
"role": "user",
"content": "What does this diagram show? Describe the data flow.",
"images": ["path/to/architecture-diagram.png"]
}
]
)
print(response["message"]["content"])
Document Understanding
Llama 3.2 Vision handles:
Scanned documents - extract text, tables, and structure from PDFs
Charts and graphs - read data values and describe trends
Screenshots - analyze UI, identify errors, extract information
Photographs - describe content, identify objects, read text
# Via HuggingFace transformers
from transformers import MllamaForConditionalGeneration, AutoProcessor
from PIL import Image
model = MllamaForConditionalGeneration.from_pretrained(
"meta-llama/Llama-3.2-11B-Vision-Instruct",
device_map="auto"
)
processor = AutoProcessor.from_pretrained("meta-llama/Llama-3.2-11B-Vision-Instruct")
image = Image.open("invoice.png")
inputs = processor(image, "Extract all line items and totals from this invoice.", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(output[0]))
On-Device With 1B and 3B
The 1B and 3B text-only models are optimized for mobile via ExecuTorch (Meta's mobile inference framework). They fit in 500MB-1.5GB of device memory - practical for iOS and Android applications that need local inference without a network call.
Summary
Llama 3.2 brings competitive vision capability to the open-source ecosystem. The 11B vision model is particularly compelling: single-GPU, commercially licensed, and matching GPT-4o mini on MMMU. Get the weights at HuggingFace and read the release post at Meta AI.
Practical deep-dives on LLMs, developer tools, and AI engineering. No filler. Unsubscribe any time.
// written byFIG. AUTH-01
530
Mahmudul Haque Qudrati
CEO & ML Engineer
CEO and ML Engineer at Pristren. Builds AI-powered software for teams and writes about machine learning, LLMs, developer tools, and practical AI applications.
OpenCode vs Claude Code: Open-Source Agentic CLI Compared
OpenCode runs Claude, GPT, Gemini, or local Ollama models in one terminal agent — Claude Code is official, polished, and Anthropic-native. Honest 2026 comparison.
