This article was originally published on Best GPU for LLM. The full version with interactive tools, FAQ, and live pricing is on the original site.
Llama 4 is Meta's most capable open model yet — and its Mixture-of-Experts architecture makes it more accessible than the parameter count suggests. The RTX 5090 is the best GPU for Llama 4 Scout locally, with 32GB VRAM fitting the full Q4 quantization. Maverick is a different story: at 400B total parameters, it requires multi-GPU or cloud.
See the recommended pick on the original guide
Understanding Llama 4's MoE architecture
Llama 4's Mixture-of-Experts design is the key to understanding its hardware requirements. Unlike dense models where every parameter activates for every token, MoE models route each token through only a subset of "expert" layers.
- Scout (109B total, 17B active) — 109B parameters exist in memory, but only 17B activate per token. Inference speed resembles a 17B dense model, but you still need VRAM to hold all 109B weights.
- Maverick (400B total, 17B active) — Same 17B active parameter count as Scout, but a much larger expert pool. Requires ~80GB VRAM at Q4 — beyond any single consumer GPU.
This is the critical distinction: active parameters determine speed, total parameters determine VRAM requirements. Scout does not need 109B worth of computation per token, but it does need 109B worth of memory.
At Q4_K_M, Scout's 109B weights compress to ~25GB — just within the RTX 5090's 32GB capacity.
Scout vs Maverick: which should you target?
| Feature | Scout (109B) | Maverick (400B) |
|---|---|---|
| Active params per token | 17B | 17B |
| Total model weights (FP16) | ~218GB | ~800GB |
| Q4_K_M size | ~25GB | ~80GB |
| Minimum VRAM | 32GB | 80GB+ |
| Single consumer GPU? | RTX 5090 only | No |
| Benchmark quality | > Llama 3 70B | > Llama 3 405B |
| Best local option | RTX 5090 | RunPod cloud |
Scout is the practically deployable model. Maverick is what you access when you need the best quality and are willing to use cloud inference — see our Llama 4 Maverick hardware summary for the full breakdown of what Maverick requires. If you are specifically evaluating Scout as a standalone target, our Llama 4 Scout GPU guide goes deeper on its inference characteristics and hardware recommendations.
VRAM requirements by model and quantization
| Model | Quantization | VRAM Required | Fits On |
|---|---|---|---|
| Scout (109B) | Q2_K | ~14GB | RTX 4090, RTX 5080, RTX 4060 Ti 16GB |
| Scout (109B) | Q3_K_M | ~19GB | RTX 4090, RTX 5090 |
| Scout (109B) | Q4_K_M | ~25GB | RTX 5090 (32GB) |
| Scout (109B) | Q6_K | ~35GB | 2x RTX 4090 (48GB), RTX 6000 Ada |
| Scout (109B) | Q8 | ~45GB | 2x RTX 4090 |
| Maverick (400B) | Q4_K_M | ~80GB | 4x RTX 4090, data center GPUs |
| Maverick (400B) | Q8 | ~150GB | Data center only |
VRAM chart available at the original article
GPU benchmarks for Llama 4 Scout
Benchmarks via llama.cpp at Q4_K_M, single-GPU unless noted:
| GPU | VRAM | Scout Q4_K_M | Scout Q3_K_M | Price |
|---|---|---|---|---|
| RTX 5090 | 32GB | ~18 tok/s | ~22 tok/s | ~$2,000 |
| 2x RTX 4090 | 48GB | ~12 tok/s | ~15 tok/s | ~$3,200 |
| RTX 4090 | 24GB | Q4 won't fit | ~20 tok/s | ~$1,600 |
| RTX 3090 | 24GB | Q4 won't fit | ~16 tok/s | ~$900 used |
Note: The RTX 4090's 24GB falls short by ~1GB for Scout at Q4_K_M. Q3_K_M at ~19GB fits, but quality drops noticeably. The RTX 5090's 32GB is the comfortable single-card solution.
See the recommended pick on the original guide
Dual-GPU setup guide for Llama 4 Scout
If you already own an RTX 4090 or want maximum quality without upgrading to RTX 5090, a dual-GPU setup via llama.cpp tensor splitting is the answer:
What you need:
- 2x RTX 4090 (or any combination totaling 48GB+)
- llama.cpp with CUDA support compiled
- NVLink is not required — PCIe works fine for inference
Running Scout on dual 4090s:
# Compile llama.cpp with CUDA
make LLAMA_CUDA=1
# Run Scout with tensor split across two GPUs
./llama-cli -m scout-q4_k_m.gguf \
--n-gpu-layers 999 \
--tensor-split 0.5,0.5 \
-p "Your prompt here"
The --tensor-split 0.5,0.5 flag distributes layers evenly across both GPUs. llama.cpp handles the inter-GPU communication automatically over PCIe — no NVLink required for inference.
Throughput on dual 4090s is lower than a single RTX 5090 (~12 vs ~18 tok/s) because PCIe bandwidth becomes a bottleneck versus the 5090's internal memory bandwidth. But dual 4090s can run Scout at Q6_K (~35GB) where the 5090 cannot, offering better output quality at the cost of speed.
KV cache and long context requirements
Llama 4 Scout supports very long context windows. The KV cache VRAM cost scales with context length:
| Context Length | KV Cache Size (Scout) | Total VRAM (Q4_K_M) |
|---|---|---|
| 4K tokens | ~1GB | ~26GB |
| 16K tokens | ~4GB | ~29GB |
| 32K tokens | ~8GB | ~33GB |
| 64K tokens | ~16GB | ~41GB |
At 32K context, Scout at Q4_K_M no longer fits on a 32GB RTX 5090. For long context workloads, use Q3_K_M to free up headroom, or step down to 16K context maximum. Dual RTX 4090s (48GB) give more flexibility for long context at Q4_K_M.
Maverick on cloud: RunPod
For Maverick-class workloads, cloud is the practical answer. RunPod provides A100 80GB and H100 instances that handle Maverick at Q4 without multi-GPU setup complexity.
A single A100 80GB fits Maverick at Q4_K_M (~80GB) at close margins. For comfortable Maverick inference with context headroom, an H100 80GB or a pair of A100s is ideal.
How Llama 4 compares to previous generations
Understanding the hardware shift from Llama 3 helps set expectations:
| Model | VRAM at Q4_K_M | Single GPU? | Best Single GPU |
|---|---|---|---|
| Llama 3 8B | ~5GB | Yes (any) | RTX 3060 12GB |
| Llama 3 70B | ~40GB | No | Needs dual GPU |
| Llama 4 Scout | ~25GB | Yes (RTX 5090) | RTX 5090 |
| Llama 4 Maverick | ~80GB | No | Cloud only |
Scout is the first time a Meta frontier model fits on a single consumer GPU at Q4 quality. This is a significant milestone — the MoE architecture delivers Llama 3 70B-beating performance in a package that a $2,000 consumer card can handle.
Which GPU should you buy for Llama 4?
Running Scout at best quality? → RTX 5090 ($2,000). The only consumer GPU that fits Scout at Q4_K_M. 32GB VRAM, ~18 tok/s, up to 16K context with headroom.
Already own an RTX 4090? → Add a second RTX 4090 ($1,600). Two 4090s give 48GB combined — Scout at Q6_K for better quality than even the 5090 at Q4. If you plan to use Scout in a retrieval-augmented setup, the second card pays for itself fast through KV-cache headroom alone.
On a budget with a 24GB card? → Use Q3_K_M. Scout at Q3_K_M fits an RTX 4090 or 3090 at ~19GB. Quality is reduced but workable for most tasks.
Need Maverick? → RunPod cloud. No single consumer GPU handles Maverick. RunPod A100 80GB instances are the most cost-effective option.
Not sure which Llama 4 size to target? → Start with Scout on RTX 5090. Scout already beats Llama 3 70B benchmarks — Maverick is overkill for most local use cases.
Common mistakes to avoid
- Buying a 24GB card expecting Scout to fit at Q4_K_M — it does not. You need 32GB. Use Q3_K_M on 24GB cards or upgrade to the RTX 5090.
- Treating total parameter count as VRAM requirement — Scout has 109B total parameters but only 17B activate per token. VRAM need is set by all weights, but inference speed reflects only active parameters.
- Attempting Maverick on a single consumer GPU — even the RTX 5090's 32GB falls far short of the 80GB minimum for Maverick at Q4. Use cloud.
- Ignoring KV cache growth for long-context use — Scout at Q4_K_M leaves only ~7GB headroom on an RTX 5090. Long context (32K+) overflows VRAM. Use Q3_K_M or cap context if you need long conversations.
- Comparing Scout directly to Llama 3 70B as equivalent — Scout beats Llama 3 70B on most benchmarks. The full expert pool provides qualitative improvements beyond what the active parameter count suggests. If reasoning quality is your priority and Scout's MoE behavior feels uneven on long chains, our DeepSeek GPU guide covers an alternative dense-reasoning family that runs comfortably on the same 24-32GB hardware tier.
Final verdict
| Goal | Best GPU | Price |
|---|---|---|
| Llama 4 Scout (Q4, best quality) | RTX 5090 | ~$2,000 |
| Llama 4 Scout (Q6, dual GPU) | 2x RTX 4090 | ~$3,200 |
| Llama 4 Scout (Q3, budget option) | RTX 4090 | ~$1,600 |
| Llama 4 Maverick | RunPod cloud | Pay per hour |
See the recommended pick on the original guide
Llama 4's MoE design makes Scout the most capable model to ever fit on a single consumer GPU. The RTX 5090 is the unlock — no other single card gets you there at usable quality.
For full VRAM breakdowns at every quantization level, see how much VRAM for Llama 4. If you are coming from an older setup, the Llama 3 GPU guide covers the previous generation. Comparing Llama 4 against Alibaba's competing release? See our best GPU for Qwen 3 guide for the dense alternative. For VRAM sizing fundamentals, the local LLM VRAM guide explains the math.
Read the full guide on Best GPU for LLM — includes our VRAM calculator, GPU comparison table, and live pricing.
Top comments (0)