Quick answer: Llama 4 Scout needs ~67GB at Q4_K_M and Maverick ~245GB. Neither is a single-consumer-GPU model — not even on a 32GB RTX 5090. The MoE architecture keeps active compute low, which is why these models feel fast, but all expert weights still load into VRAM and that is what you have to buy for.
Why Llama 4’s VRAM math is unusual
Llama 4 is a Mixture-of-Experts (MoE) model. Both Scout and Maverick have 17B active parameters — meaning only 17B parameters compute per token. But the full set of expert weights still lives in VRAM. You cannot load only the active experts because routing decisions happen at inference time.
- Scout (109B total, 17B active): Full weight loading requires fitting 109B params in VRAM, not 17B
- Maverick (400B total, 17B active): Same principle at much larger scale
This is why the VRAM requirement looks disproportionate to the “active” size.
Llama 4 Scout VRAM requirements
Scout has 109B parameters across 16 experts.
| Quantization | VRAM Required | Fits On | Quality |
|---|---|---|---|
| FP16 (full) | 217GB | Data center / multi-GPU | Reference quality |
| Q8_0 | 117GB | Multi-GPU (2x A100 80GB) | Near-lossless |
| Q4_K_M | 67GB | A100/H100 80GB, or 3x RTX 5090 | Good — the practical build |
Scout’s three figures are the published downloads for llama4:scout, checked 2026-09-13 — 217GB, 117GB and 67GB respectively, so they are exact rather than estimated. Maverick’s Q4_K_M below is published too, at 245GB; its FP16 and Q8_0 rows are derived from the parameter count and keep a tilde.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Those are the only three quantizations Ollama publishes for Scout, and the sizes
are its own: llama4:17b-scout-16e-instruct-q4_K_M is a 67GB download.
The key threshold is one nobody wants to hear: there is no consumer card in this conversation. Q4_K_M at ~67GB is more than double the RTX 5090’s 32GB and nearly triple the RTX 4090’s 24GB. A dual-5090 build at 64GB is still short. The realistic local configurations are three RTX 5090s, a pair of 48GB workstation cards, or a single 80GB A100/H100 — and at that point cloud rental is usually the cheaper answer for anything short of constant use.
If you came here because “17B active parameters” sounded like a 24GB model, that is the whole trap: 17B is what computes per token, 109B is what has to be resident.
Llama 4 Maverick VRAM requirements
Maverick has 400B parameters across 128 experts.
| Quantization | VRAM Required | Fits On | Quality |
|---|---|---|---|
| FP16 (full) | ~803GB | Large cluster | Reference quality |
| Q8_0 | ~428GB | Large cluster | Near-lossless |
| Q4_K_M | 245GB | 4x A100 80GB or equivalent | Good |
Practical takeaway: Maverick is not a local inference model for most people. Even at aggressive quantization, it requires multi-GPU setups that cost $3,000-10,000+. Cloud inference via RunPod or similar is the pragmatic path. For a full hardware cost analysis, see our Llama 4 Maverick hardware summary.
GPU recommendations by VRAM tier
| VRAM | Best Option | Llama 4 Scout | Llama 4 Maverick |
|---|---|---|---|
| 24GB | RTX 4090 / RTX 3090 | Not possible | Not possible |
| 32GB | RTX 5090 | Not possible | Not possible |
| 48GB | 2x RTX 4090 | Not possible | Not possible |
| 80GB | A100 / H100, or 3x RTX 5090 | Q4_K_M | Not possible |
| 320GB+ | 4x A100 80GB | Q8_0 | Q4_K_M |
NVIDIA GeForce RTX 5090
32GB GDDR7At 32GB it takes three to hold Scout's 67GB at Q4. Worth pricing against cloud before you commit — and consider whether a 32B dense model does the job.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
How KV cache affects VRAM
The numbers above are base model weights only. KV cache — which stores attention states for context — adds VRAM on top:
- 4K context: ~1-2GB additional
- 8K context: ~2-4GB additional
- 32K context: ~8-12GB additional
- 128K context (Scout’s maximum): ~30-50GB additional
These sit on top of the 67GB of base weights, which is why an 80GB A100 is the smallest sensible home for Scout rather than a card that merely matches the weight size: 80GB leaves ~13GB for cache, good for roughly 32K context. Full 128K context on Scout is a multi-card proposition at any quantization.
Why there is no budget quantization for Scout
For a dense model, a card that is too small is usually solved by dropping a quantization level — Q3 or Q2 buys back a third of the footprint at some cost in quality. That escape hatch does not exist here. Ollama publishes exactly three builds for Llama 4 (Q4_K_M, Q8_0, FP16), and the smallest is 67GB.
Community GGUF quantizations below Q4 do circulate, but a 109B MoE degrades badly under aggressive quantization: the router’s decisions get noisier as the expert weights lose precision, so the failure mode is not “slightly worse prose” but visibly wrong expert selection. Below Q4 you are usually better served by a well-quantized dense model that fits your card properly — Qwen 3 32B at ~20GB is the obvious comparison on a 24GB card.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→Common mistakes to avoid
- Treating active parameter count as the VRAM requirement. Scout activates 17B parameters per token and needs ~67GB of VRAM. All expert weights must be resident even though only 17B compute per forward pass — the gap between those two numbers is the single most expensive misconception about MoE models.
- Hunting for a sub-Q4 build to make Scout fit. Ollama publishes none, and community quantizations below Q4 degrade a 109B MoE badly because the router loses precision along with the experts.
- Attempting Maverick on a single GPU of any kind. At ~245GB for Q4_K_M it needs roughly four 80GB cards. Even a single H100 is short by a factor of three. This is a cloud-only model.
- Ignoring context window needs. KV cache sits on top of model weights. On an 80GB card holding Scout’s 67GB, the remaining ~13GB is good for about 32K context — the 128K the model supports is a multi-card proposition.
Final verdict table
| Setup | Model | Quantization | Verdict |
|---|---|---|---|
| Cloud A100/H100 80GB | Scout | Q4_K_M | The sensible default — rent before you build |
| 4x RTX 3090 used (96GB) | Scout | Q4_K_M | Cheapest local build, ~$3,280 |
| 3x RTX 5090 (96GB) | Scout | Q4_K_M | Same capacity, ~$14,700 |
| Any single consumer card | Scout | — | Does not fit at any published build |
| Cloud multi-GPU (4x 80GB) | Maverick | Q4_K_M | The only option |
Llama 4 Scout is the most capable model ever designed to fit on a single consumer GPU — but only just. The RTX 5090’s 32GB clears the Q4 bar by 7GB. No other single card does.
For GPU recommendations specifically for Llama 4, see the best GPU for Llama 4 buyer’s guide. If you are focused solely on Scout as your deployment target, the Llama 4 Scout GPU guide covers its hardware requirements in dedicated detail. For general VRAM sizing across all models, the local LLM VRAM guide covers the full framework. And for the previous generation comparison, the Llama 70B GPU guide provides context on how hardware requirements have evolved.