Vision-languageMixture-of-Experts109B params17B active10M max context

Llama 4 Scout 17B-16E (MoE, VLM) VRAM Requirements: How Many GPUs & What It Costs

Llama 4 Scout 17B-16E (MoE, VLM) needs 60.0 GB of VRAM for its weights at AWQ 4-bit — 60.8 GB once a KV cache is loaded. Below is the full breakdown: which GPUs fit, how many you need, what it serves, and what it costs per million tokens. Every figure comes from tokcalc's open-source capacity model.

Weights (AWQ 4-bit)
60.0 GB
60.8 GB with KV cache
Fits on 1 GPU?
Yes
H100 SXM5 80GB or larger
Cheapest config
$0.0146
8× RTX 3080 10GB / 1M tokens
Fastest config
49920
tok/s on 2× Cerebras CS-3 (WSE-3)
VRAM requirement by quantization
Weight footprint = params × bytes per param. The “with KV cache” column adds one request at 4K context.
QuantizationWeightsWith KV cacheSmallest single GPU
FP16218.0 GB218.8 GBB300 SXM 288GB (288 GB)
FP8 (E4M3)109.0 GB109.8 GBIntel Gaudi 3 128GB (128 GB)
AWQ 4-bitrecommended60.0 GB60.8 GBMac M3 Max (64GB) (64 GB)
Which GPU should you use?
Reference workload: 8K context, 8 concurrent users, continuous batching, AWQ 4-bit. Ranked by fitting first, then by tokens per dollar-hour.
GPUGPUsTotal VRAMDecode tok/sAggregate tok/sCost / 1M tokens
RTX 3080 10GB8×80 GB316.230351$0.0146
RTX 4080 16GB8×128 GB298.328634$0.0233
RTX 5090 32GB4×128 GB372.717891$0.0311
RTX 3090 24GB4×96 GB194.78376$0.0332
RTX A4000 16GB8×128 GB186.417891$0.0373
RTX 4090 24GB4×96 GB209.710064$0.0442
T4 16GB8×128 GB133.17624$0.0583
V100 SXM2 32GB4×128 GB187.28986$0.0618
A10G 24GB4×96 GB124.85990$0.0668
TPU v6e Trillium (32GB)4×128 GB170.48177$0.1087
A100 40GB SXM42×80 GB161.23869$0.1292
L4 24GB4×96 GB62.402995$0.1484

RTX 3080 10GB has no NVLink, so 8-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.

How concurrency changes with context length
On 8× RTX 3080 10GB at AWQ 4-bit. KV cache scales linearly with context, so long contexts cut the number of users you can serve from the same GPUs.
ContextConcurrent usersKV per requestFits?
4K24824.6 MB✓ fits
8K121.61 GB✓ fits
33K36.44 GB✓ fits
66K112.88 GB✓ fits
131K025.77 GB✗ does not fit
Frequently asked questions

Can Llama 4 Scout 17B-16E (MoE, VLM) run on a single GPU?

Yes. At AWQ 4-bit the model needs 60.8 GB of VRAM, so it fits on a single H100 SXM5 80GB or anything with more VRAM. At full FP16 it needs 218.8 GB, which fits on one B300 SXM 288GB or larger.

How many GPUs does Llama 4 Scout 17B-16E (MoE, VLM) need?

At AWQ 4-bit with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 3080 10GB (80 GB total VRAM). $0.0146 per 1M output tokens.

How fast can Llama 4 Scout 17B-16E (MoE, VLM) be served?

On 2× Cerebras CS-3 (WSE-3) at AWQ 4-bit, the aggregate throughput is 49920 tokens/sec with continuous batching. That hardware has no published rental price, so its cost per million tokens is not comparable.

How much VRAM does Llama 4 Scout 17B-16E (MoE, VLM) need?

Weights alone: 218.0 GB at FP16, 60.0 GB at AWQ 4-bit. Add KV cache for your context length and concurrency — every concurrent request at 8,192 tokens costs extra VRAM, so budget headroom above the weight figure.

Should I run Llama 4 Scout 17B-16E (MoE, VLM) in FP8 or FP16?

FP8 halves the weight footprint (218 GB → 109 GB) and is faster than FP16 on Hopper and Blackwell, where it has native kernels. FP16 remains the reference quality bar; FP8 loses a small amount of accuracy in exchange for roughly doubling the batch size you can serve per GPU.

Why does Llama 4 Scout 17B-16E (MoE, VLM) need more VRAM than its active parameter count suggests?

Llama 4 Scout 17B-16E (MoE, VLM) is a mixture-of-experts model: 17B parameters are active per token, but all 109B experts must stay resident in VRAM. Capacity is bounded by the resident set, while decode throughput is bounded by the active set — so it serves fast per token but needs the full footprint.

Is it cheaper to self-host Llama 4 Scout 17B-16E (MoE, VLM) than to call an API?

At 8× RTX 3080 10GB the marginal cost is $0.0146 per 1M output tokens, before you add idle capacity, ops staff, or egress. That is typically cheaper than frontier API pricing at sustained volume, but API calls win when your load is bursty or too small to keep a GPU busy. The Build vs Buy tab on tokcalc prices this against a specific provider for you.

Try Llama 4 Scout 17B-16E (MoE, VLM) with your own context and concurrency

The calculator uses the exact same formulas as this page — change the workload and watch every number update.

Open in calculator

Llama 4 Scout 17B-16E (MoE, VLM) on specific GPUs

Llama 4 Scout 17B-16E (MoE, VLM) on RTX 5090 32GBLlama 4 Scout 17B-16E (MoE, VLM) on AMD MI300X 192GBLlama 4 Scout 17B-16E (MoE, VLM) on L40S 48GBLlama 4 Scout 17B-16E (MoE, VLM) on AMD MI325X 288GBLlama 4 Scout 17B-16E (MoE, VLM) on A100 80GB SXM4Llama 4 Scout 17B-16E (MoE, VLM) on Intel Gaudi 3 128GBLlama 4 Scout 17B-16E (MoE, VLM) on H100 SXM5 80GBLlama 4 Scout 17B-16E (MoE, VLM) on B200 SXM 192GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters109.00B
Active parameters per token17.00B
ArchitectureMixture-of-Experts
Attention layoutFull attention (GQA)
Transformer layers48
Hidden dimension5120
Attention heads40 query / 8 KV (GQA)
Head dimension128
Vocabulary size202,000
Max context (catalog)10M
KV cache per token @ 8K192.0 KB

Every layer caches keys and values for the whole context, so the KV cache grows linearly with conversation length.