Run Llama 4 Locally: Scout and Maverick Setup With Ollama

Can you run Llama 4 locally? Yes, but it is a big-memory job. Llama 4 Scout is a 67GB download in Ollama (llama4:scout, Q4_K_M) and Maverick is 245GB (llama4:maverick). Both are Mixture-of-Experts models with only 17B parameters active per token, which keeps them fast for their size, but every expert must be loaded, so you need memory for the whole file: a 96-128GB Mac or Strix Halo machine for Scout, a 256GB+ workstation for Maverick. On a 16-24GB GPU with 64GB+ of system RAM, llama.cpp can run Scout by keeping the experts in RAM.

What's New in Llama 4

Meta released Llama 4 on April 5, 2025 (Llama 4 model card). The headline changes:

Key Features

Model Variants

Llama 4 Scout - The one most people can actually run

• 109B total parameters, 17B active, 16 experts
• 10M-token context window
• 67GB download at Q4_K_M in Ollama
• MoE: no tok/s figure published here (speed depends on how experts are placed)

Llama 4 Maverick - The stronger model

• 400B total parameters, 17B active, 128 experts
• 1M-token context window
• 245GB download at Q4_K_M in Ollama
• MoE: no tok/s figure published here

Llama 4 Behemoth - The teacher model

• About 2T total parameters, 288B active
• Previewed by Meta as the teacher for Scout and Maverick
• Weights not released
• Not something you can run locally

To see how 17B active parameters can sit inside a 400B model, read our Mixture of Experts explainer.

Local Setup Guide

Step 1: Install Ollama

# macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows - download from ollama.com

Step 2: Pull Llama 4

# Scout (67GB download; needs ~80GB+ of RAM/VRAM/unified memory)
ollama pull llama4:scout

# Maverick (245GB download; workstation/server class)
ollama pull llama4:maverick

The explicit tags are llama4:17b-scout-16e-instruct-q4_K_M and llama4:17b-maverick-128e-instruct-q4_K_M; Q8_0 and FP16 builds are also listed (sizes below).

Step 3: Run and Test

ollama run llama4:scout

Test multimodal:

>>> Describe this image: /path/to/image.jpg

Then check where the model loaded:

ollama ps
# PROCESSOR shows how much is on GPU vs CPU

Step 4: Configure for Performance

Create a custom Modelfile:

cat > Llama4Modelfile << 'EOF'
FROM llama4:scout

# Keep the context modest: the KV cache comes on top of 67GB of weights
PARAMETER num_ctx 8192
PARAMETER temperature 0.7
PARAMETER top_p 0.9

SYSTEM "You are Llama 4, a helpful AI assistant with vision capabilities."
EOF

ollama create llama4-custom -f Llama4Modelfile

Smaller GPU, lots of RAM? llama.cpp can keep the MoE expert weights in system RAM while attention runs on the GPU:

# Keep all expert weights on the CPU side
llama-server -m Llama-4-Scout-17B-16E-Instruct-Q4_K_M.gguf -ngl 99 --cpu-moe -c 8192

# Or only the experts of the first N layers, if part of them fits in VRAM
llama-server -m Llama-4-Scout-17B-16E-Instruct-Q4_K_M.gguf -ngl 99 --n-cpu-moe 30 -c 8192

Hardware Requirements

VRAM Requirements

Download sizes from Ollama's llama4 tags (registry, November 2, 2026). Add a few GB for context.

Recommended Configurations

Apple Silicon

Using Vision Capabilities

Image Analysis

import ollama

response = ollama.chat(
    model='llama4:scout',
    messages=[{
        'role': 'user',
        'content': 'What do you see in this image?',
        'images': ['./chart.png']
    }]
)
print(response['message']['content'])

Document Processing

# Analyze a PDF page as image
from pdf2image import convert_from_path
import ollama

pages = convert_from_path('document.pdf')
pages[0].save('page1.png', 'PNG')

response = ollama.chat(
    model='llama4:scout',
    messages=[{
        'role': 'user',
        'content': 'Extract and summarize the key information from this document.',
        'images': ['page1.png']
    }]
)

Code Understanding from Screenshots

response = ollama.chat(
    model='llama4:scout',
    messages=[{
        'role': 'user',
        'content': 'Explain what this code does and identify any bugs.',
        'images': ['code_screenshot.png']
    }]
)

Benchmark Results

All figures below are Meta's, from the Llama 4 model card (instruction-tuned models, 0-shot). We have not run them ourselves. Meta compares against its own earlier models, so that is the comparison shown.

Language Understanding

Coding

Vision

Integration Examples

LangChain Integration

from langchain_ollama import ChatOllama

llm = ChatOllama(
    model="llama4:scout",
    temperature=0.7
)

response = llm.invoke("Explain quantum computing simply")
print(response.content)

Open WebUI Setup

docker run -d -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  ghcr.io/open-webui/open-webui:main

Access http://localhost:3000 and select llama4:scout.

Llama 4 vs Competition

Key Takeaways

1. Llama 4 brings image input to Meta's open-weight models
2. MoE architecture keeps compute per token at 17B, but memory must hold the full 109B or 400B
3. Scout is the realistic local model: 67GB at Q4, for 96-128GB machines or a GPU plus lots of RAM
4. Maverick is workstation/server class at 245GB
5. Vision benchmarks are strong on Meta's own card (DocVQA 94.4, ChartQA 90.0 for Maverick)

Next Steps

1. Compare to DeepSeek R1 for reasoning tasks
2. Understand MoE memory vs speed before buying hardware
3. Size Llama 4 for your machine with the MoE calculator
4. Build multimodal agents with Llama 4
5. Set up RAG with vision-enhanced retrieval
6. Pick hardware for large models

---

Llama 4 puts a strong multimodal model in open weights. With enough memory, Scout runs on a single high-capacity machine; for everyone else, its MoE design is the reason a GPU-plus-RAM setup can still run it.