Knowledge Base · Local AI
How much VRAM to run Llama 4 locally?
By Tiny · how I pick
As an Amazon Associate I earn from qualifying purchases. The hardware links below are affiliate links.
Llama 4 Scout needs about 55 to 60 GB of memory at 4-bit. It's a 109B-parameter Mixture-of-Experts model, and the whole thing has to sit in memory even though only 17B of it works on each word. That puts it on a Mac with 64 GB or more, or a rig with two or three graphics cards. A single consumer card can't run it.
Llama 4 Maverick, at about 400B, is a data center model. On a normal graphics card, run something else.
Words you'll see in this guide
- VRAM: the memory on your graphics card. AI models load into it, so more VRAM means you can run bigger models.
- Mixture-of-Experts (MoE): a model design that only uses part of itself for each word it writes, so it runs faster than its total size suggests. The whole model still has to be loaded.
- Parameters (17B, 109B): a rough measure of a model's size. "8B" means 8 billion. Bigger models are usually smarter and need more memory.
- Unified memory: Apple's design where the processor and graphics share one fast pool of memory, so a Mac's RAM works like VRAM.
Llama 4 memory at a glance
| Model | Total size | Memory (4-bit) | Realistic home hardware |
|---|---|---|---|
| Llama 4 Scout | 109B (17B active) | About 55 to 60 GB | A Mac with 64 GB or more, or two or three 24 GB cards |
| Llama 4 Maverick | About 400B | 200 GB or more | A multi-card server, not a home setup |
What to run on a normal graphics card instead
With a typical 8 to 24 GB card, a model sized for your hardware will give you better answers, faster:
- 16 GB card: Qwen 3 14B, Gemma 4 12B or the DeepSeek-R1 14B distill.
- 24 GB card: Qwen 3 32B, which keeps up with much larger models on code and reasoning. A used RTX 3090 is the usual 24 GB card.
- Mac with 64 GB or more: this is where Llama 4 Scout runs well, because of unified memory. A high-memory Mac Studio is the simplest way to get there. See Mac vs PC for local AI.
How to run it
On hardware that can handle it, the simple way is Ollama with Open WebUI. Check your machine against the model with the Server Calculator first, so you don't download 60 GB for nothing.
Where to go next
Quick answers
How much VRAM do I need to run Llama 4 Scout?
Around 55 to 60GB at 4-bit. It only uses 17B parameters at a time, but it has 109B in total, and all of them must sit in memory. That's beyond any single regular graphics card, so you need a 64GB or larger Apple Silicon Mac, or two to three 24GB cards.
Why does Llama 4 need so much memory if only 17B is active?
It's a mixture-of-experts model that picks which parts to use as it goes, so all 109B parameters must be loaded even though only about 17B work on each word. Memory depends on the total, not the active part.
Can I run Llama 4 Maverick at home?
For almost everyone, no. It has around 400B parameters and needs well over 200GB of memory even compressed.
What should I run instead of Llama 4 on a normal GPU?
A model sized for your card: Qwen 3 14B or a DeepSeek-R1 14B distill on 16GB, or Qwen 3 32B on 24GB. They run fast and give excellent results in the memory you have.
Free email updates
Meta and others keep shrinking these models. Subscribe and I'll tell you when a Llama 4-class model fits on one card.
No spam. Unsubscribe anytime. Privacy policy.