VRAMGlass contains affiliate links. As an eBay Partner, we may be compensated if you make a purchase.As an eBay Partner, we may be compensated.Learn more

GPUs tracked57Price history49 daysPrice points (24h)43,400 pointsBiggest 24h dropGeForce RTX 3060 12GB-11.1%Biggest 24h riseCMP 170HX 40GB+19.8%Updated
$ USD

How to Deploy Llama 4 Scout 17B 16E Locally

Llama 4 Scout is a multimodal MoE that Meta released in April 2025: 108.6B parameters, 17B used per token, with image understanding. Meta says it supports 10 million tokens of context, but actually filling that would take more than 500 GB; what you can realistically open locally is a few hundred thousand to just over a million. At 4-bit with 32K it takes about 69 GB, so the 96GB RTX PRO 6000 or a 128GB unified-memory machine holds it entirely.

Llama 4 Scout 17B 16E details

  • Min VRAM at 4-bit69 GB · 8K
  • Parameters108.6B
  • Active params17B
  • Max context10M
Architecture
MoE
Experts offloaded
11 GB VRAM + 59 GB RAM
License
Llama 4 Community License
Launch date
Apr 5, 2025
Vendor
Meta AI
Hugging Face repo
meta-llama/Llama-4-Scout-17B-16E-Instruct
GGUF repo
unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF
NVFP4 repo (Blackwell)
nvidia/Llama-4-Scout-17B-16E-Instruct-NVFP4
MLX repo (Apple)
mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit
Ollama tag
llama4:scout
Layers
48
Hidden size
5,120
KV heads
8
Head dim
128

How much VRAM does Llama 4 Scout 17B 16E need for local deployment?

QuantizationWeightsGB32K contextGB128K contextGB256K contextGB10M context (max)GB
1-bit35GGUF38.943.750.2552.7
2-bit42.4GGUF46.351.157.6560.1
3-bit49GGUF52.957.764.2566.7
4-bit65.4GGUF69.374.180.6583.1
5-bit76.5GGUF80.485.291.7594.2
8-bit114.5GGUF118.4123.2129.7632.2
BF16/FP16215.6GGUF219.5224.3230.8733.3

Full calculation rules are on the Methodology page.

Can a single GPU run Llama 4 Scout 17B 16E?

No consumer card can; 4-bit takes a workstation or data center card. The RTX PRO 6000, the 80GB A100 and the 72GB RTX PRO 5000 all hold 4-bit whole with plenty of context, while the 64GB MI210 just squeezes it in with only about 20K of context left. A 48GB RTX A6000 has to drop to 2-bit, and a 32GB RTX 5090 can't fit even that. Once it fits, though, it's quick. Each token only passes through a small slice of the model's experts, about 17B parameters' worth, so the RTX PRO 6000 does about 79 tokens/s. The same trait helps unified-memory machines: the 96GB Mac Studio M5 Ultra manages about 61 tokens/s at 4-bit, and 128GB boxes like the DGX Spark still a dozen-plus, far better than a dense model this size. With only a 24GB card, you can keep the experts in system RAM instead.

eBay
DevicePriceSpeed t/sMax context
Ryzen AI Max+ 395 128GBWhole system · 128 GB$4,46420 listings16.41.4M
Instinct MI21064 GB$4,99411 listings74.220K
Mac Studio M5 Ultra 96GBWhole system · 96 GB$7,27412 listings61703K
MacBook Pro M5 Max 128GBWhole system · 128 GB$7,54920 listings35.61.4M
DGX Spark 128GBWhole system · 128 GB$8,24022 listings17.41.4M
RTX PRO 5000 72GB72 GB$10,7223 listings64.9191K
A100 80GB PCIe80 GB$18,99912 listings82.4361K
RTX PRO 600096 GB$22,50011 listings78.6703K
Mac Studio M3 Ultra 512GBWhole system · 512 GB$22,90013 listings44.99.4M

Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.

How many GPUs do you need to run Llama 4 Scout 17B 16E?

At 4-bit, the smallest build is a pair of 48GB RTX A6000 cards: split across GPUs with vLLM at 32K of context, it needs about 69 GB, and the pair just covers that at about 41 tokens/s. 24GB cards (RTX 3090, 4090, RX 7900 XTX) need four, since two only add up to 48GB; with four cards' worth of memory bandwidth, they actually edge out the A6000 pair at close to 50 tokens/s. Two 32GB RTX 5090 cards only fit 2-bit, the fastest option but with noticeably weaker answers. For 8-bit, plan on two 96GB RTX PRO 6000s or four 5090s.

eBay
DeviceCardsTotal priceTotal VRAMSpeed t/sTotal power
Tesla V100 32GB32 GB per card4 cards$3,580128 GB46.41,000 W
Radeon RX 7900 XTX24 GB per card4 cards$5,00096 GB48.81,420 W
GeForce RTX 309024 GB per card4 cards$7,19696 GB47.91,400 W
Arc Pro B7032 GB per card4 cards$7,400128 GB33.8920 W
RTX A600048 GB per card2 cards$9,79896 GB40.9600 W
Instinct MI21064 GB per card2 cards$9,988128 GB71.7600 W
GeForce RTX 409024 GB per card4 cards$15,03296 GB50.71,800 W
GeForce RTX 509032 GB per card4 cards$33,580128 GB762,300 W
GeForce RTX 4090 48GB48 GB per card2 cardsNo price96 GB50.7900 W

Full calculation rules are on the Methodology page.

What PC or server build do you need to run Llama 4 Scout 17B 16E locally?

Each token goes through only 1 of 16 routed experts plus 1 shared expert, so it still runs with the experts in system RAM. At 4-bit, a used 8-channel DDR4 server with an RTX 3090 does about 26 tokens/s, a dual-channel DDR5 desktop with a 16GB card about 16, and a DDR4 desktop under 10; at 2-bit, the DDR5 desktop reaches about 26.

eBay

2 × GeForce RTX 5070 + Desktop DDR5

Home PC
Generation speed16.5t/s
Memory bandwidth96GB/s
GPU
2 × GeForce RTX 507011 / 48 expert layers in 24 GB VRAM
$1,798
RAM
64 GB DDR5-60004 × 16 GB sticks
$900
CPU
AMD Ryzen 7 77008 cores / 16 threads · 2-channel
$212
Board
Gigabyte B650 EAGLE AX4 DIMM slots
$114
PSU
Corsair RM850x850 W
$70
Reference build price$3,094

2 × Tesla V100 32GB + Dual Xeon E5 DDR4

Budget
Generation speed44.7t/s
Memory bandwidth137GB/s
GPU
2 × Tesla V100 32GB46 / 48 expert layers in 64 GB VRAM
$1,790
RAM
128 GB DDR4-21338 × 16 GB sticks
$632
CPU
2 × Intel Xeon E5-2680 v428 cores / 56 threads · 8-channel in total
$38
Board
Supermicro X10DRi16 DIMM slots
$193
PSU
Corsair RM1200x SHIFT1,200 W
$135
Reference build price$2,788

GeForce RTX 3090 + Single EPYC DDR4

Best value
Generation speed26.2t/s
Memory bandwidth171GB/s
GPU
GeForce RTX 309012 / 48 expert layers in 24 GB VRAM
$1,799
RAM
128 GB DDR4-26668 × 16 GB sticks
$944
CPU
AMD EPYC 753232 cores / 64 threads · 8-channel
$262
Board
Supermicro H12SSL-i8 DIMM slots
$1,100
PSU
Corsair RM850x850 W
$70
Reference build price$4,175

GeForce RTX 5090 + Single EPYC DDR5

Flagship
Generation speed56.6t/s
Memory bandwidth461GB/s
GPU
GeForce RTX 509019 / 48 expert layers in 32 GB VRAM
$8,395
RAM
192 GB DDR5-480012 × 16 GB sticks
$7,164
CPU
AMD EPYC 935432 cores / 64 threads · 12-channel
$1,693
Board
Supermicro H13SSL-N12 DIMM slots
$1,000
PSU
Corsair RM1200x SHIFT1,200 W
$135
Reference build price$18,387

Detailed calculation rules are on the Methodology page.

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.

What to know before running Llama 4 Scout 17B 16E locally

  1. How to enable images

    Ollama's llama4:scout supports image input out of the box. llama.cpp needs the mmproj file for images loaded alongside the model; with the GGUFs packaged by ggml-org, -hf downloads it automatically. Meta tested up to 5 images at a time.

  2. Why long context is lighter than the 70B

    In every 4 layers, 3 only attend within small 8K chunks and just 1 sees the full text, so at 4-bit, going from 32K to 256K adds only about 11 GB. On Meta's long-context test MTOB (translating from half a book up to a full book), Llama 3.3 70B has no score because its context is only 128K, while Scout does.

  3. Quantization and frameworks

    Meta released only BF16 weights and says they fit on a single H100 with on-the-fly int4 quantization; the vLLM recipe also lists NVIDIA's FP8 version. For home use, go with GGUF: 4-bit is about 65 GB and 3-bit about 49 GB. Sampling defaults are unchanged: temperature 0.6 and top_p 0.9.

  4. Languages and license

    It officially supports 12 languages, not including Chinese. The Llama 4 use policy also states that individuals living in the EU and companies with their principal place of business in the EU are not granted rights to use the multimodal models (end users of products built on them aren't restricted), and Scout is a multimodal model.

Launch commands

Ollama

  • Windows
  • macOS
  • Linux
Terminal
ollama run llama4:scout

llama.cpp

  • NVIDIA
  • AMD
  • Mac
  • CPU
All on GPU
Terminal
llama-server -hf unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M \
  -c 8192
MoE offloading

Short on VRAM? Add --cpu-moe to keep the routed experts in system RAM (about 59 GB needed).

Terminal
llama-server -hf unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M \
  -c 8192 -ngl 99 --cpu-moe

vLLM

  • Linux
  • NVIDIA
  • AMD
Original weights
Terminal
vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct
NVFP4 · Blackwell

Blackwell GPUs (RTX 50 series, RTX PRO 6000, B200) run NVFP4 weights natively, at about 30% of the original size. For multiple GPUs, add --tensor-parallel-size with the card count.

Terminal
vllm serve nvidia/Llama-4-Scout-17B-16E-Instruct-NVFP4

MLX

  • Apple Silicon

Vision models run on mlx-vlm: on a Mac, pip install mlx-vlm first; this command downloads the MLX weights from Hugging Face and starts a server.

Terminal
mlx_vlm.server --model mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology

Other models in the Meta family

ModelTotal paramsActive params4-bit size
Llama 3.2 3BDense3.2B3.2B2 GB
Llama 3.1 8BDense8B8B4.9 GB
Llama 3.3 70BDense70.6B70.6B42.5 GB
Muse Spark OpenMoE816B42B505.5 GB

About Meta