Can you run Llama 4 Scout at home, and what does its licence allow?

Meta released Llama 4 Scout on April 5 as an open-weight model: you can download the trained parameters (the "weights") and run them on your own equipment. Scout is a mixture-of-experts model, meaning it contains 16 specialist sub-networks, or experts, and uses only part of them for each token it processes. It activates 17 billion parameters at a time out of 109 billion in total. That makes it fast for its size, but it doesn't make it small: all 109 billion parameters still have to sit in memory.

How much hardware it needs

Meta's model card says Scout fits on a single H100, an 80 GB data-centre GPU, when its weights are compressed to 4 bits on the fly. That compression is called quantization: storing each parameter with fewer bits, trading a little quality for a lot of memory. A rough calculation shows what that means for home machines.

These are estimates for the weights only; the model needs extra memory for the text it is working on, and more for long inputs. In short, a typical gaming PC with a 12 or 24 GB graphics card can't hold Scout. Realistic home options are a workstation with several large GPUs, or a Mac whose unified memory (shared between processor and graphics) is big enough for a quantized version.

Software support is arriving quickly. On April 7, llama.cpp, the popular open-source program for running models on ordinary computers, added text-only Llama 4 support. Its developer tested Scout in an 8-bit format on a Mac Studio with an M3 Ultra chip and 512 GB of memory and reported about 27 tokens per second, a comfortable reading speed. Image input isn't supported there yet, so for now a local Scout setup is mostly a text assistant.

A simple test once it runs

Before you trust any local model with real work, give it a short, repeatable test. The tasks below are a suggested set; score each answer yourself from 0 to 2.

1. Summarize this 2-page document in five bullet points. [paste a document you know well]
2. Answer in French: what are three ways to reduce home heating costs?
3. Write a Python function that removes duplicate lines from a text file, keeping order.
4. Here is a table of monthly sales. Which month grew most, and by how much? [paste a small table]
5. What happened in the news last week?   (A good answer admits it doesn't know.)

Run the same five tasks on the model you use today. Question 5 matters: Scout's knowledge stops in August 2024, and a model that invents recent events is a model to double-check.

What the licence allows and requires

Llama 4 is released under the Llama 4 Community License Agreement, effective April 5, 2025. Here is a summary, not legal advice; read the full text before you build a product on it.

• Allowed: using, copying, modifying, fine-tuning and redistributing the model, royalty-free, including in commercial products.
• Attribution: if you distribute the model or a product that contains it, include a copy of the licence, show "Built with Llama" prominently (for example on your website or in your documentation), and keep the required notice file.
• Naming: if you use Llama 4 or its outputs to train or improve another model that you then distribute, that model's name must begin with "Llama."
• Very large companies: if your products counted over 700 million monthly users in the calendar month before Llama 4's release date, you must ask Meta for a separate licence, which it can refuse.
• Acceptable use: you must follow the law and Meta's Acceptable Use Policy, which the licence incorporates. That policy also withholds the licence for Llama 4's multimodal models from individuals living in, and companies based in, the European Union, though not from end users of products built with them.
• No warranty, and suing Meta over intellectual property in Llama ends your licence.

So it is open in the practical sense that matters to most builders, but it is not an open-source licence in the traditional sense, because of these conditions.

How does it compare with Llama 3?

The design is different. Llama 3 models are dense, meaning every parameter works on every token, while Llama 4 is Meta's first mixture-of-experts family. Scout also accepts images natively; the Llama 3.3 70B model in Meta's comparison has no image support. And the context window jumps from 128,000 tokens to 10 million, enough in principle for a large code base or several books at once.

On Meta's own benchmark table for the instruction-tuned models, Scout beats Llama 3.3 70B on knowledge and reasoning tests (74.3 versus 68.9 on MMLU Pro, 57.2 versus 50.5 on GPQA Diamond), while the two are about even on a coding test (32.8 versus 33.3 on LiveCodeBench) and on multilingual math (90.6 versus 91.1). These are Meta's figures, measured at full precision; quantized versions at home may score a little lower.

The practical trade-off is memory. Because all 109 billion parameters must be loaded, Scout needs more memory than the 8B and 70B Llama 3 models, in exchange for images, a far longer context and somewhat better reasoning. For most home users with modest hardware, a smaller Llama 3 model remains the practical choice until lighter Llama 4 options or better quantizations appear.