Skip to content

Best GPU for Llama 4 Scout: VRAM Requirements

Llama 4 Scout needs about 241.6 GB of GPU memory in BF16 and 61.7 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.

Llama 4 Scout needs about 241.6 GB of VRAM in FP16, 121.7 GB in FP8 and 61.7 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on one NVIDIA B300 288GB; a 4-bit build fits on one NVIDIA A100 PCIe 80GB.

Llama 4 Scout is a mixture-of-experts language model with 109B total and 17B active parameters per token from Meta, released in April 2025, with a 10,485,760-token context window and published under the Llama 4 Community License. Its weights alone take 218 GB in BF16 and 54.5 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.

Recommended GPUs for Llama 4 Scout

Llama 4 Scout VRAM requirements by precision

One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.

PrecisionWeightsKV cacheTotal for inference
BF16 / FP16218 GB1.6 GB241.6 GB
FP8109 GB1.6 GB121.7 GB
INT454.5 GB1.6 GB61.7 GB

Llama 4 Scout VRAM by task

Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.

TaskGPU memory
Inference at the full 10,485,760-token context (FP8)2,387.6 GB
QLoRA (4-bit base)70.9 GB
LoRA (BF16 base)283.4 GB
Full fine-tuning (BF16, AdamW)1,918.4 GB

GPUs that fit Llama 4 Scout

Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.

GPUMemoryBF16 inference (241.6 GB)FP8 inference (121.7 GB)INT4 inference (61.7 GB)QLoRA (70.9 GB)LoRA (283.4 GB)
NVIDIA GeForce RTX 5080 16GB16 GB—8 GPUs4 GPUs8 GPUs—
NVIDIA GeForce RTX 4090 24GB24 GB—8 GPUs4 GPUs4 GPUs—
NVIDIA L4 24GB24 GB—8 GPUs4 GPUs4 GPUs—
NVIDIA GeForce RTX 5090 32GB32 GB8 GPUs4 GPUs2 GPUs4 GPUs—
NVIDIA Tesla V100 32GB32 GB8 GPUs4 GPUs2 GPUs4 GPUs—
NVIDIA A100 PCIe 40GB40 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA L40S 48GB48 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA Quadro RTX 8000 48GB48 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA RTX 6000 Ada 48GB48 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA RTX A6000 48GB48 GB8 GPUs4 GPUs2 GPUs2 GPUs8 GPUs
NVIDIA A100 PCIe 80GB80 GB4 GPUs2 GPUs1 GPU1 GPU4 GPUs
NVIDIA A100 SXM 80GB80 GB4 GPUs2 GPUs1 GPU1 GPU4 GPUs
NVIDIA H100 PCIe 80GB80 GB4 GPUs2 GPUs1 GPU1 GPU4 GPUs
NVIDIA H100 SXM 80GB80 GB4 GPUs2 GPUs1 GPU1 GPU4 GPUs
NVIDIA RTX PRO 6000 Blackwell 96GB96 GB4 GPUs2 GPUs1 GPU1 GPU4 GPUs
NVIDIA H200 141GB141 GB2 GPUs1 GPU1 GPU1 GPU4 GPUs
NVIDIA B200 180GB180 GB2 GPUs1 GPU1 GPU1 GPU2 GPUs
NVIDIA B300 288GB288 GB1 GPU1 GPU1 GPU1 GPU1 GPU
NVIDIA GB300 NVL288 GB1 GPU1 GPU1 GPU1 GPU1 GPU

What is Llama 4 Scout?

Llama 4 Scout is a natively multimodal mixture-of-experts model from Meta with 16 experts and 17 billion active parameters out of 109 billion. It reads text and images, writes text and supports a context window of up to 10 million tokens, the longest of the Llama 4 family.

Developer
Meta
Released
April 2025
Parameters
109B total, 17B active per token
Context window
10,485,760 tokens
License
Llama 4 Community License

Three of every four layers use chunked attention over 8,192 tokens, so the real KV cache at long context is far below this estimate, which counts every layer at full length. Meta states that Scout fits on a single H100 GPU with on-the-fly INT4 quantization.

Source: Llama 4 Scout 17B 16E Instruct model card (as of November 17, 2026).

Frequently asked questions

How much VRAM does Llama 4 Scout need?
For inference with an 8,192-token context and one request, Llama 4 Scout needs about 241.6 GB in BF16, 121.7 GB in FP8, 61.7 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
Can Llama 4 Scout run on a single GPU?
Yes. In BF16 it needs about 241.6 GB, which fits one NVIDIA B300 288GB (288 GB).
How much GPU memory does fine-tuning Llama 4 Scout take?
QLoRA (4-bit base): about 70.9 GB; LoRA (BF16 base): about 283.4 GB; Full fine-tuning (BF16, AdamW): about 1,918.4 GB. Longer training sequences and larger batches add activation memory on top.
What is the best GPU for Llama 4 Scout?
Best for inference: 1 GPU × NVIDIA H200 141GB (FP8 or INT8 weights, 121.7 GB at 8,192 tokens). Best for fine-tuning: 2 GPUs × NVIDIA B200 180GB (LoRA on BF16 weights, 283.4 GB). Budget pick: 1 GPU × NVIDIA RTX PRO 6000 Blackwell 96GB (Smallest card class that holds a 4-bit build, 61.7 GB). Check live offers before booking, as availability changes.
Can Llama 4 Scout run on an RTX 4090?
Not on one card: even a 4-bit build needs about 61.7 GB, so it takes 4 RTX 4090 cards (96 GB combined) with the model split across them.
Can Llama 4 Scout run on an RTX 5090?
Not on one card: even a 4-bit build needs about 61.7 GB, so it takes 2 RTX 5090 cards (64 GB combined) with the model split across them.

Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.

Explore the live market

Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.