Scout is mixture-of-experts, so you're buying for 109B in memory, not 17B active. The honest GPU ladder — why a 32GB card is the sweet spot, when 24GB works, and when unified memory beats them all.
The question I get about Llama 4 Scout is always "which GPU do I need," and the honest answer starts with a correction: you're not buying for a 17B model, you're buying for a 109B one. Scout is mixture-of-experts, so even though only 17B parameters fire per token, the whole 109B has to sit in memory. That single fact reshapes the whole buying decision — this is a VRAM problem first and a compute problem a distant second. So let me give you the honest hardware ladder for Scout, from the card that makes it comfortable to the one that barely gets you in the door.
Scout is a VRAM decision first: even at 17B active, all 109B parameters must fit in memory. · Illustration by Aliteq / generated with Higgsfield
The honest hardware ladder for Scout
Running Llama 4 Scout by hardware tier
Unified memory (96–128GB)
Quality Q4 (~62GB)?
Yes, comfortably
How
Full Q4_K_M + big context
Verdict
The best home pick — capacity + value
80GB card (H100 — data-center)
Quality Q4 (~62GB)?
Yes
How
Q4 + long context
Verdict
Overkill/rental; not a home buy
Multi-GPU (2× 32GB / 48GB pro)
Quality Q4 (~62GB)?
Yes, with tuning
How
Split across cards
Verdict
Works, but pricey vs unified memory
RTX 5090 (32GB) single
Quality Q4 (~62GB)?
No — low-bit only
How
~1.78-bit dynamic (~33GB), fast
Verdict
Speed pick, quality compromised
RTX 4090 / 3090 (24GB)
Quality Q4 (~62GB)?
No
How
Dynamic quant + CPU offload, ~20 tok/s
Verdict
Tinkering only
16GB and below
Quality Q4 (~62GB)?
No
How
—
Verdict
Run a smaller dense model instead
Quality Q4 (~62GB)?
How
Verdict
Unified memory (96–128GB)
Yes, comfortably
Full Q4_K_M + big context
The best home pick — capacity + value
80GB card (H100 — data-center)
Yes
Q4 + long context
Overkill/rental; not a home buy
Multi-GPU (2× 32GB / 48GB pro)
Yes, with tuning
Split across cards
Works, but pricey vs unified memory
RTX 5090 (32GB) single
No — low-bit only
~1.78-bit dynamic (~33GB), fast
Speed pick, quality compromised
RTX 4090 / 3090 (24GB)
No
Dynamic quant + CPU offload, ~20 tok/s
Tinkering only
16GB and below
No
—
Run a smaller dense model instead
If you take one line from this: for a quality Scout at home, buy memory capacity, and unified memory gives you the most of it per dollar. A fast 32GB GPU is great if you're happy with a low-bit quant for speed, but it can't hold the quality quant — and low-bit quantization chips away at exactly the qualities (long-context fidelity, output quality) that would be your reason to run Llama 4 in the first place.
What I'd actually buy
If you want a quality Scout at home: unified memory — a 96–128GB Strix Halo or Mac Studio holds the full Q4 with context headroom, and it's usually the cheapest route to that capacity.
If you want peak token speed and accept a low-bit quant: a 32GB RTX 5090. It runs Scout fast at an aggressive dynamic quant — just know you're trading quality for speed.
If you already own a 24GB card: don't rush to upgrade — run Scout with dynamic quantization plus CPU expert-offload to experiment, and lean on a smaller dense model (Qwen3, Gemma) for daily driving.
If it's occasional: rent an 80GB card by the hour rather than buying — a full Q4 Scout on rented silicon is cheap for a one-off.
Quick answers
Can an RTX 4090 run Llama 4 Scout?
Only with aggressive ~1.78-bit dynamic quantization plus CPU expert-offload, at around 20 tokens/second — usable for experimentation, not quality work. Even a 32GB RTX 5090 can only hold a low-bit quant of Scout; a quality Q4 (~62GB) needs unified memory or a 64GB+ card.
How much VRAM do I need for Llama 4 Scout?
For a quality Q4_K_M quant, ~62GB — because MoE loads all 109B — plus KV-cache headroom (up to ~24GB at 128k context). That means a 64GB+ card or 96–128GB unified memory. A 32GB card fits only an aggressive low-bit quant.
Is a 16GB GPU enough?
Not meaningfully for Scout — the 109B MoE doesn't fit well even heavily quantized, and what does fit is badly compromised. On 16GB you're far better off with a strong smaller dense model like Qwen3 or Gemma.
GPU or unified memory for Llama 4?
For a quality Scout, unified memory (96–128GB Strix Halo or Mac Studio) is the better pick — it holds the full Q4 MoE plus context, often for less than stacking GPUs. A 32GB GPU wins only on raw token speed, and only at a lower-quality low-bit quant.
