Vision-languageMixture-of-Experts400B params17B active1M max context

Llama 4 Maverick 17B-128E (MoE, VLM) VRAM Requirements: How Many GPUs & What It Costs

Llama 4 Maverick 17B-128E (MoE, VLM) needs 220.0 GB of VRAM for its weights at AWQ 4-bit — 220.8 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)
220.0 GB
220.8 GB with KV cache
Fits on 1 GPU?
Yes
B300 SXM 288GB or larger
Cheapest config
$0.0155
8× RTX 5090 32GB / 1M tokens
Fastest config
798720
tok/s on 8× 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
FP16800.0 GB800.8 GB— needs multi-GPU
FP8 (E4M3)400.0 GB400.8 GB— needs multi-GPU
AWQ 4-bitrecommended220.0 GB220.8 GBB300 SXM 288GB (288 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 5090 32GB8×256 GB745.571565$0.0155
V100 SXM2 32GB8×256 GB374.435942$0.0309
A100 40GB SXM48×320 GB644.861901$0.0323
L40S 48GB8×384 GB359.434505$0.0451
RTX 6000 Ada 48GB8×384 GB399.438339$0.0464
TPU v6e Trillium (32GB)8×256 GB340.732708$0.0544
A100 80GB SXM44×320 GB416.019968$0.0668
H100 SXM5 80GB4×320 GB696.833446$0.0831
RTX PRO 6000 Blackwell 96GB4×384 GB372.717891$0.0932
AMD MI300X 192GB2×384 GB551.213229$0.1050
H100 PCIe 80GB4×320 GB416.019968$0.1113
Intel Gaudi 3 128GB2×256 GB343.28237$0.1349

RTX 5090 32GB 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 5090 32GB 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?
4K44824.6 MB✓ fits
8K221.61 GB✓ fits
33K56.44 GB✓ fits
66K212.88 GB✓ fits
131K125.77 GB✓ fits
Frequently asked questions

Can Llama 4 Maverick 17B-128E (MoE, VLM) run on a single GPU?

Yes. At AWQ 4-bit the model needs 220.8 GB of VRAM, so it fits on a single B300 SXM 288GB or anything with more VRAM.

How many GPUs does Llama 4 Maverick 17B-128E (MoE, VLM) need?

At AWQ 4-bit with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 5090 32GB (256 GB total VRAM). $0.0155 per 1M output tokens.

How fast can Llama 4 Maverick 17B-128E (MoE, VLM) be served?

On 8× Cerebras CS-3 (WSE-3) at AWQ 4-bit, the aggregate throughput is 798720 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 Maverick 17B-128E (MoE, VLM) need?

Weights alone: 800.0 GB at FP16, 220.0 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 Maverick 17B-128E (MoE, VLM) in FP8 or FP16?

FP8 halves the weight footprint (800 GB → 400 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 Maverick 17B-128E (MoE, VLM) need more VRAM than its active parameter count suggests?

Llama 4 Maverick 17B-128E (MoE, VLM) is a mixture-of-experts model: 17B parameters are active per token, but all 400B 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 Maverick 17B-128E (MoE, VLM) than to call an API?

At 8× RTX 5090 32GB the marginal cost is $0.0155 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 Maverick 17B-128E (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 Maverick 17B-128E (MoE, VLM) on specific GPUs

Llama 4 Maverick 17B-128E (MoE, VLM) on A100 80GB SXM4Llama 4 Maverick 17B-128E (MoE, VLM) on H100 SXM5 80GBLlama 4 Maverick 17B-128E (MoE, VLM) on B200 SXM 192GBLlama 4 Maverick 17B-128E (MoE, VLM) on H200 SXM5 141GBLlama 4 Maverick 17B-128E (MoE, VLM) on AMD MI325X 288GBLlama 4 Maverick 17B-128E (MoE, VLM) on B300 SXM 288GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters400.00B
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)1M
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.