Llama 4 Scout is Meta's model for people who want a capable LLM on their own hardware. Our AMI puts it on an AWS GPU instance behind an OpenAI-compatible API, so your data stays in your VPC and your existing OpenAI code keeps working after a base-URL change.
This page covers what Scout is, what it takes to run, what its license actually lets you do, and where it earns its keep.
In short
What Is Llama 4 Scout?
LLaMA 4 Scout is part of Meta's latest generation of open-weight models, released in April 2025. It's a Mixture-of-Experts model: 109 billion parameters in total, split across 16 experts, with about 17 billion active for any given token. That's why it answers faster than a dense model of the same size, and also why it still needs a lot of GPU memory, since every expert has to be loaded.
- Mixture-of-Experts, 16 experts, 17B active and 109B total parameters
- Text and image input, text output
- A context window Meta advertises at up to 10 million tokens
- Released under the Llama 4 Community License (more on that below)
The full specs are in Meta's Llama 4 model card. For how Scout stacks up against Qwen, DeepSeek and the rest, see our roundup of the best open source LLMs to self-host.
Llama 4 Scout API: what the AMI gives you
The AMI exposes an API in the OpenAI format: /v1/chat/completions,
completions, embeddings and a model list. Point the OpenAI SDK, LangChain or
any tool that speaks that format at your instance, change the model name to
llama4-scout, and it runs against your own server instead.
We explain why that matters in self-hosting an OpenAI-compatible API on AWS.
What's included in the AMI
- Ubuntu 22.04 with the NVIDIA drivers, CUDA and Python stack already installed
- The model downloaded and served behind the OpenAI-compatible API
- A CloudFormation template that launches the server and prints the API documentation URL in its Outputs tab
- Automatic SSL for the domain you point at the server, with manual steps if it fails
Is Llama 4 Scout multimodal?
The model is: Meta trained it on text and images together, so it can read a screenshot, a chart or a scanned form and answer in text. The AMI's endpoints are the OpenAI text ones listed above. If image input is the reason you're here, launch an instance and test it with your own files before you design a product around it.
Hardware: what Llama 4 Scout needs on AWS
Only 17 billion parameters are active per token, but all 109 billion have to sit in memory. Our Marketplace listing runs on g4dn.metal, a bare-metal instance with eight NVIDIA T4 GPUs (128 GB of GPU memory in total), 96 vCPUs and 384 GB of RAM.
That's a big instance, and you pay for it by the hour. Stop it outside working hours if it serves an internal team, and check your account's GPU vCPU quota before launching, because new AWS accounts often can't start a g4dn.metal until AWS raises the limit.
Llama 4 Scout use cases
| Use Case | Industry Examples | What LLaMA 4 Scout Can Do |
|---|---|---|
| Customer support with screenshots | E-commerce, SaaS, Fintech | Agents that understand screenshots, bills, or photo uploads |
| Document review | Healthcare, Pharma, HR | Summarize and check documents against a policy, with a person signing off |
| Tutoring | EdTech, Nonprofits | Explain graphs, diagrams, or annotated content |
| Accessibility | Media, Design, Web Agencies | Draft alt text, captions, and content summaries for visual media |
| Coding help | DevOps, Software Agencies | Explain architecture diagrams, UI mockups, or screenshots of code errors |
| Long-document Q&A | Legal, Research, Consulting | Answer questions across long contracts, reports or codebases in one prompt |
The common thread is data you'd rather not send to a hosted API: contracts, patient documents, internal code. That, more than raw benchmark scores, is the reason to run Scout yourself.
The license: what "open" means for Llama 4
Llama 4 is open-weight, not open source in the OSI sense. The Llama 4 Community License lets you run, fine-tune and build on Scout commercially, with conditions:
- Products built on it must show "Built with Llama".
- Services with more than 700 million monthly active users need a separate license from Meta.
- Individuals and companies based in the EU aren't granted rights to the multimodal Llama 4 models, though EU users can still use a product built on them by a non-EU company. Meta's Llama FAQ explains the detail.
If you're in the EU, read that last point with your lawyer before you build. It's the single biggest difference between Llama 4 and models under Apache 2.0.
Meetrix AMI vs doing it yourself
| Feature | Meetrix AMI | DIY Deployment |
|---|---|---|
| Setup Time | A CloudFormation launch plus the model start-up | Hours or days of driver, CUDA and serving setup |
| GPU Support | Drivers and CUDA pre-installed and tested | Manual CUDA/drivers required |
| API | OpenAI-compatible endpoints out of the box | You choose and configure a serving stack |
| SSL | Set up automatically for your domain | Manual setup needed |
| Support | Meetrix support by email | None or community only |
Doing it yourself is perfectly reasonable if you already run vLLM or Ollama; our Ollama vs vLLM comparison will help you pick. If you'd rather start from a working server, the Llama 4 Scout AMI on AWS Marketplace is the shortcut, and the Llama 4 Scout store page lists the guide and video in one place.
Frequently Asked Questions
Does the Llama 4 Scout AMI have an API?
Yes. It serves an OpenAI-compatible API with chat, completions, embeddings and model-listing endpoints, so code written for the OpenAI SDK works by changing the base URL and model name. The developer guide has curl examples for each endpoint.
Is Llama 4 Scout multimodal?
The model is. Meta trained Scout to take text and images together and answer in text. Our AMI exposes the OpenAI-style text endpoints, so if image input matters to your project, test it on your own instance before you build around it.
Is Llama 4 Scout open source?
It's open-weight, not open source. You can download and run it, but under Meta's Llama 4 Community License: you must show "Built with Llama", services over 700 million monthly users need Meta's permission, and EU-based companies don't get rights to the multimodal models.
What AWS instance does Llama 4 Scout need?
Our Marketplace listing runs on g4dn.metal, which has eight NVIDIA T4 GPUs with 128 GB of GPU memory in total. Scout has 109 billion parameters overall, even though only 17 billion are active per token, so the whole model still has to fit in memory.
How much does the Meetrix Llama 4 Scout AMI cost?
The software fee on AWS Marketplace is $0.048 per hour on g4dn.metal, billed on top of the EC2 instance itself. The GPU instance is most of the bill, so stop it when you're not using it.