Quick answer: No single consumer GPU runs Llama 4 Scout at Q4_K_M, where it is a 67GB download — more than double the RTX 5090’s 32GB. Third-party GGUFs go lower: Q2_K is 39.6GB and fits two 24GB cards, at a real quality cost. The practical local builds are three RTX 5090s or four used RTX 3090s; for most people renting an 80GB A100 or H100 is both cheaper and faster to set up.
NVIDIA GeForce RTX 5090
32GB GDDR732GB each, and Scout's Q4 weights are 67GB — so this is a three-card build, not a one-card one. Price it against an hourly A100 first.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What is Llama 4 Scout?
Llama 4 Scout is Meta’s mixture-of-experts (MoE) model released in April 2025. The architecture:
- Total parameters: 109 billion
- Active parameters per forward pass: 17 billion (only 16 experts are active at once)
- Context window: up to 10 million tokens (theoretical)
- Architecture: 16 experts, 1 active per token
The MoE design means Scout behaves like a 17B dense model during inference — not a 109B model. You only compute 17B worth of activations per token. But you must store all 109B parameters in memory simultaneously, which is the VRAM bottleneck.
Llama 4 Scout VRAM requirements
| Quantization | Weight Size | VRAM Needed | Single Card? |
|---|---|---|---|
| FP16 | ~217GB | 220GB+ | No — multi-GPU cluster |
| Q8_0 | ~117GB | 125GB+ | No |
| Q4_K_M | ~67GB | 80GB+ | No — 3x RTX 5090 or an 80GB card |
Note on MoE quantization: MoE models do not compress better than dense ones.
Every expert block is quantized like any other weight, and every one of them has
to be in memory because routing is decided per token at inference time. Scout’s
109B parameters at roughly 4.5 bits each come to ~67GB, which is exactly what
llama4:17b-scout-16e-instruct-q4_K_M downloads. The three rows above are the
only quantizations Ollama publishes for this model.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
GPU benchmarks for Llama 4 Scout
Estimated performance running Scout Q4_K_M with llama.cpp:
| GPU | VRAM | Fits Scout Q4? | ~Tok/s | Price |
|---|---|---|---|---|
| RTX 5090 | 32GB | No — 35GB short | — | ~$4,900 |
| RTX 4090 | 24GB | No | — | ~$2,200 |
| 2x RTX 3090 | 48GB | No — 19GB short | — | ~$1,640 used |
| 2x RTX 5090 | 64GB | No — 3GB short | — | ~$9,800 |
| 3x RTX 5090 | 96GB | Yes, with context room | ~40 tok/s | ~$14,700 |
| 4x RTX 3090 | 96GB | Yes | ~25 tok/s | ~$3,280 used |
| Cloud A100 80GB | 80GB | Yes, comfortable | ~45 tok/s | RunPod ~$2/hr |
| Cloud H100 80GB | 80GB | Yes, very fast | ~80 tok/s | RunPod ~$3.5/hr |
Two numbers there are worth sitting with. A dual RTX 5090 build — nearly $10,000 of graphics cards — comes three gigabytes short of loading this model at Q4. And four used RTX 3090s, at roughly a third of that price, clear it with room to spare. If you are set on running Scout locally, the used 24GB market is the only build that makes financial sense; everything else points at renting.
Check RTX 3090 Price (for a four-card build)→Buy on Shopee SG→Best GPU options for Llama 4 Scout
Cloud 80GB — the default answer
A single A100 or H100 with 80GB holds Scout’s 67GB of Q4 weights with roughly 13GB spare. Runtime overhead takes about 1.5GB of that, leaving 11.5GB for cache — around 60K context at 0.19MB per token. RunPod rents A100s around $2/hr. Against a four-card local build at ~$3,280 that is about 1,600 hours of use before buying wins — well over a year at four hours a day. For most people evaluating Scout, this is where the story should end.
4x RTX 3090 used (~$3,280) — cheapest local build
Four used 24GB cards give 96GB via tensor parallelism in llama.cpp, which clears Scout’s 67GB with real context headroom. It is by a wide margin the cheapest hardware that runs this model, and it is the only consumer build that makes financial sense. The trade is everything around the cards: a motherboard with four usable PCIe slots, roughly 1,400W of power headroom, and a case and cooling plan that takes the problem seriously.
3x RTX 5090 (~$14,700) — the same capacity, four times the price
Also 96GB, considerably faster per card, and hard to justify unless you already own the cards or need the throughput for something other than Scout. Note that two 5090s do not work: 64GB is three gigabytes short of the Q4 weights, so the jump is straight from two cards that fail to three that succeed.
A single card — and what the low-bit builds change
Scout at Ollama’s Q4_K_M is 67GB against 32GB on the largest consumer card, so one GPU does not run the build most people mean. But Ollama is not the only source, and that changes the answer for anyone who already owns two cards. Unsloth publishes GGUF quantizations of Scout well below Q4; these are their download sizes, checked 2026-09-13:
| Build | Size | Smallest hardware that holds it |
|---|---|---|
| UD-TQ1_0 | 29.3GB | One RTX 5090 (32GB), with almost nothing left for context |
| UD-IQ1_S | 32.5GB | Half a gigabyte past a 5090 — so two cards |
| Q2_K | 39.6GB | 2x RTX 3090 or 2x RTX 4090 (48GB) |
| Q3_K_S | 46.7GB | 2x 24GB, with no context headroom |
| UD-Q3_K_XL | 49GB | 2x RTX 5090 (64GB) |
| Q4_K_M | 67GB | 96GB of consumer cards, or one 80GB data-centre card |
So the accurate answer for a dual-3090 owner is not “it will not fit” — Q2_K fits with room for context. The question is whether you want it to. Scout is sixteen experts, each one far smaller than a dense 109B model’s layers and each quantized on its own, so quality falls off faster below 3 bits here than it does on a dense model of the same file size. The comparison to weigh is not Scout at Q2 against Scout at Q4; it is Scout at Q2 against Qwen 3 32B at Q4, which costs 20GB and fits a single card. Low-bit Scout earns its place if you need Scout specifically — its long-context design, or its behaviour on your prompts. It is not a cheaper road to general quality.
Run Llama 4 Scout on RunPod A100→ Rent H100 on Vast.ai for Scout inference→Which setup should YOU use for Llama 4 Scout?
- Just want to try Scout? Cloud GPU on RunPod. An 80GB A100 at ~$2/hr runs it comfortably today, with nothing to buy and nothing to assemble.
- Want it locally at the lowest price? 4x RTX 3090 used (~$3,280 total). 96GB combined, and the only consumer configuration that both fits the model and makes financial sense. Budget for the motherboard and power supply too.
- Already own an RTX 5090? For Q4_K_M you need two more — 32GB and 64GB both fall short of 67GB, and 96GB is the first configuration that clears it. A second 5090 does open Unsloth’s 49GB Q3_K_XL.
- Have an RTX 4090 or 3090 and hoped Scout would fit? One card, no. Two give you 48GB, which holds Q2_K at 39.6GB — but ask first whether Qwen 3 32B (~20GB) at Q4 on the card you already own does the job better.
Common mistakes to avoid
- Thinking Scout runs like a 17B model in terms of VRAM. The 17B active parameter count is a compute figure, not a memory figure. All 109B worth of expert weights are resident at once, which is why the Q4_K_M download is 67GB rather than 11GB. The exception worth knowing is an architecture that genuinely executes in stages — Nemotron TwoTower calls its two towers alternately, so one can sit in CPU memory while the other runs.
- Buying any single card specifically for Scout. Not the 4090, not the 5090. The smallest thing that runs it is an 80GB data-centre card or four consumer 24GB cards.
- Assuming
ollama pullis the whole picture. Ollama publishes Q4_K_M, Q8_0 and FP16 for Llama 4 and nothing else, so there is no official Q3 or Q2 to fall back on — but third-party GGUFs do go down to 29.3GB, which is what a dual-card owner should be looking at rather than concluding the model is out of reach. - Underestimating context overhead. Even at 80GB, the 67GB of weights leave 13GB spare, and once runtime overhead is counted about 11.5GB is usable as cache — around 60K context at FP16, or roughly double that with a Q8 cache. Scout is designed for far longer than that, so one card still does not exercise it.
Final verdict
| Setup | VRAM | Scout Q4_K_M? | ~Tok/s | Cost |
|---|---|---|---|---|
| RunPod A100 | 80GB | Yes | ~45 tok/s | ~$2/hr |
| 4x RTX 3090 (used) | 96GB | Yes | ~25 tok/s | ~$3,280 |
| 3x RTX 5090 | 96GB | Yes | ~40 tok/s | ~$14,700 |
| 2x RTX 5090 | 64GB | No — 3GB short | — | ~$9,800 |
| RTX 5090 | 32GB | No | — | ~$4,900 |
| RTX 4090 | 24GB | No | — | ~$2,200 |
NVIDIA GeForce RTX 3090
24GB GDDR6XFour used RTX 3090s give 96GB combined — the only consumer configuration that fits Scout's 67GB and still makes financial sense against renting.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
For the full Llama 4 lineup across all model sizes, see our best GPU for Llama 4 guide. Curious about Scout’s VRAM needs in more depth? See how much VRAM for Llama 4. For a dedicated look at what the larger Maverick model demands, see our Llama 4 Maverick hardware summary. For cloud-first LLM inference setup, check best GPU for Ollama.