How to Install and Run LLaMA 4 (Scout and Maverick) Locally on Any Operating System

In this guide, we’ll walk through the installation process for the Llama Force Scout (or any Llama 4 model) on your local machine (Windows / Linux / Mac).
Scout is an incredibly powerful model featuring a 10 million token context window and 17 billion active parameters — perfect for advanced multi-modal tasks.

Minimum PC Requirements for LLaMA 4

✅ Works across Windows, Linux, and Mac as long as:

• You have a Python environment (like virtualenv or conda).
• You have Git installed.
• You install the correct version of PyTorch for your system.

Pro Tip:
Always run installations using Windows Terminal or PowerShell — not the default Command Prompt.
Ensure both Python and Git are added to your system’s PATH.

Step 1: Set Up a Python Virtual Environment

Using Conda (recommended for simplicity):

conda create -n ai python=3.11 -y
conda activate ai

• Creates a new environment named ai with Python 3.11.
• Activates it immediately after creation.

✅ This works on Windows, Linux, and Mac (assuming Conda is installed).

Step 2: Install Required Libraries

Inside the new environment, install the key packages:

pip install torch
pip install git+https://github.com/huggingface/transformers
pip install git+https://github.com/huggingface/accelerate
pip install huggingface_hub

This will install:

• torch: Core deep learning library.
• transformers: Hugging Face library for model loading.
• accelerate: Hardware optimization for training/inference.
• huggingface_hub: Access to models and datasets from Hugging Face.

Step 3: Authenticate with Hugging Face

1. Create a Hugging Face account if you don’t already have one.
2. Get your Access Token from huggingface.co.
3. Log in through your terminal:

huggingface-cli login

You’ll be prompted to paste your token. This will allow you to access gated models like Llama 4 Scout.

Step 4: Install and Launch Jupyter Notebook

Install Jupyter Notebook and interactive widgets inside your Conda environment:

conda install -c conda-forge --override-channels notebook -y
conda install -c conda-forge --override-channels ipywidgets -y

Then start Jupyter Notebook:

jupyter notebook

• Launches Jupyter inside your ai environment.
• Now you can create a new notebook and run your model code!

Step 5: Load the Llama Force Scout Model

Inside a new Jupyter notebook, paste the following code:

from transformers import AutoProcessor, Llama4ForConditionalGeneration
import torch

# Define the model ID
model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct"

# Load the processor
processor = AutoProcessor.from_pretrained(model_id)

# Load the model
model = Llama4ForConditionalGeneration.from_pretrained(
    model_id,
    attn_implementation="flex_attention",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

✅ This code:

• Loads the tokenizer and processor.
• Loads the Scout model.
• Automatically maps to GPU if available.
• Loads weights in bfloat16 (ideal for modern GPUs like H100).

Step 6: Using the Model – Example: Compare Two Images

You can use Scout to reason over multiple images. Here’s a simple example:

# Image URLs
url1 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/0052a70beed5bf71b92610a43a52df6d286cd5f3/diffusers/rabbit.jpg"
url2 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/datasets/cat_style_layout.png"

# Define chat messages
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": url1},
            {"type": "image", "url": url2},
            {"type": "text", "text": "Can you describe how these two images are similar, and how they differ?"},
        ],
    }
]

# Process the inputs
inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

# Generate the model output
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
)

# Decode and print the output
generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(generated_text)

✅ This script:

• Fetches two images from URLs.
• Packages them with a user question.
• Sends them through Scout.
• Decodes and prints Scout’s intelligent response.

What Llama Force Scout Can Do

Advanced/Upcoming Features

• Video processing: (Coming Soon!)
• Multi-turn conversations: Maintain dialogue context over multiple messages.
• Custom Finetuning: Finetune Scout with your own datasets (requires Hugging Face + Accelerate).

Final Notes

• Always replace "model_id" with the correct Hugging Face model repo or local path.
• Make sure your GPU and CUDA drivers are properly set up for best performance.
• Scout is multi-modal: it understands both images and text, deeply and intelligently.