Llama 3 and Llama 4: Meta's open family — which model runs on your hardware

If in June 2024 Llama 3 looked like the endpoint of open models, today Meta's family is the main road to serious local AI: three generations in a year and a half, an architecture change (MoE), a jump from an 8K context to 10M tokens, and a license that means one thing for businesses — you can use it in-house for free, asking for nothing.

This article is the map of the Llama family for people who have to actually use it: what's there, what runs on your hardware, and which one to pick.

The family, one line per model

Model	Release	Architecture	Active parameters	Context	Notes
Llama 3.2 1B / 3B	Sep 2024	dense	1B / 3B	128K	on-device, embedded assistants
Llama 3.1 8B	Jul 2024	dense	8B	128K	the 8 GB workhorse
Llama 3.3 70B	Dec 2024	dense	70B	128K	the "working" 70B
Llama 4 Scout (17B×16E)	Apr 2025	MoE	17B of 109B	10M	multimodal (text+image)
Llama 4 Maverick (17B×128E)	Apr 2025	MoE	17B of 400B	1M	top tier, not for desktop

Two things jump out. First: Scout is a MoE — for each token only a subset of experts works (17 billion active parameters out of 109 total), so it is fast like a 17B and capable well beyond a dense 70B, but it occupies RAM like a 109B. Second: context became a feature — 10M tokens on Scout means, in theory, an entire archive of documents in play.

What runs on your hardware

The rule is the same as always (the full math): quantized weights + KV cache, with headroom for the OS. In practice:

Model (Q4)	RAM/VRAM required	On what hardware
Llama 3.2 3B	4–6 GB	any laptop, iGPUs, up to a Raspberry Pi 5 16GB (slow)
Llama 3.1 8B	8–10 GB	12GB GPU or a Mac with 16GB
Llama 3.3 70B	44–50 GB	2×24GB, Strix Halo 128GB, Mac with 64GB+
Llama 4 Scout Q4	~67 GB (Q4_K_M)	2×24GB barely, Strix Halo 128GB, Mac 96GB+

The non-obvious point: Scout is not a first-computer model, but it is the model that changes the mind of whoever has 128 GB. With the Mac Studio M5 Ultra or a Strix Halo mini-PC, 109B MoE at 17B active means 17B speed with budget-frontier quality — the best quality/speed per euro of RAM that exists locally today. For 109B though, 800 GB/s bandwidth is the bottleneck: below 300 GB/s (typical dual 24GB) you drop to 5–8 tok/s, usable but not brilliant.

How to use it

From Ollama, two commands for the workhorse and one for Scout:

ollama pull llama3.1:8b      # the everyday 8B
ollama pull llama4:scout     # ~67GB, needs dedicated RAM

ollama run llama3.1:8b "Summarize this contract in 5 points"

From code, it's the same OpenAI-compatible API as always (port 11434):

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
r = client.chat.completions.create(
    model="llama3.1:8b",
    messages=[{"role": "user", "content": "List the obligations in article 30"}])
print(r.choices[0].message.content)

If the app already uses the OpenAI or Anthropic SDK, change one line and your Meta models run in-house: Ollama vs LM Studio explains which runtime to pick.

The license, translated for the Italian business

The Llama 3 family ships under the Llama 3 Community License; Llama 4 under the Llama 4 Community License. For anyone who is not Meta:

• commercial use: free, including the sold product;
• the real constraint: above 700 million monthly active users a separate agreement with Meta is required — a threshold no Italian business will come close to touching;
• the brand stays: "Llama" cannot become the name of your product;
• not open source in the strict sense (not OSI-approved), but for business use it is de facto open.

For the firm, the accountant, the insurer that wants a model over internal documents: a license without surprises. What changes for GDPR is covered in the dedicated article — in short: with Llama in-house, the extra-EU transfer disappears.

What to pick today

• First computer, 16GB → Llama 3.1 8B: the best quality/speed trade-off under 128K.
• 24–48GB budget (2×GPU or Mac) → Llama 3.3 70B for text, or Scout if vision interests you.
• 128GB+ (Strix Halo, Mac Studio) → Scout: the model to show a client.
• Embedded/edge → 3.2 3B, the 1B for tighter devices.
• Just curious, no RAM → the 8B on a Raspberry Pi 5 still does its job.

The Llama family is no longer "Meta's model": it is the reference scale on which everything else is measured, and the fact that the 3.1 8B of 2024 and the 4 Scout of 2025 run on the same ollama pull line tells you how much the curve has straightened.

Sources

• Meta AI — The Llama 4 herd
• Llama 4 Community License Agreement
• Hugging Face — Llama-4-Scout-17B-16E
• Ollama library — llama4

Article generated with AI