TextMixture-of-Experts30B params3B active41K max context

Qwen 3 30B-A3B (MoE) VRAM Requirements: How Many GPUs & What It Costs

Qwen 3 30B-A3B (MoE) needs 30.0 GB of VRAM for its weights at FP8 (E4M3) — 30.4 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 (FP8 (E4M3))
30.0 GB
30.4 GB with KV cache
Fits on 1 GPU?
Yes
A100 40GB SXM4 or larger
Cheapest config
$0.005513
4× RTX 3080 10GB / 1M tokens
Fastest config
78000
tok/s on 1× 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
FP1660.0 GB60.4 GBMac M3 Max (64GB) (64 GB)
FP8 (E4M3)recommended30.0 GB30.4 GBV100 SXM2 32GB (32 GB)
AWQ 4-bit16.5 GB16.9 GBA10G 24GB (24 GB)
Which GPU should you use?
Reference workload: 8K context, 8 concurrent users, continuous batching, FP8 (E4M3). Ranked by fitting first, then by tokens per dollar-hour.
GPUGPUsTotal VRAMDecode tok/sAggregate tok/sCost / 1M tokens
RTX 3080 10GB4×40 GB839.840310$0.005513
RTX 4080 16GB4×64 GB792.338030$0.008765
RTX 5090 32GB2×64 GB990.123762$0.0117
RTX 3090 24GB2×48 GB517.111867$0.0117
RTX A4000 16GB4×64 GB495.023762$0.0140
RTX 4090 24GB2×48 GB556.913366$0.0166
T4 16GB4×64 GB353.610800$0.0206
V100 SXM2 32GB2×64 GB497.311934$0.0233
A10G 24GB2×48 GB331.57956$0.0251
AMD MI300X 192GB1×192 GB172320670$0.0336
TPU v6e Trillium (32GB)2×64 GB452.510860$0.0409
A100 40GB SXM41×40 GB503.86045$0.0414

RTX 3080 10GB has no NVLink, so 4-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 4× RTX 3080 10GB at FP8 (E4M3). 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?
4K24412.3 MB✓ fits
8K12824.6 MB✓ fits
33K33.22 GB✓ fits
66K16.44 GB✓ fits
131K012.88 GB✗ does not fit
Frequently asked questions

Can Qwen 3 30B-A3B (MoE) run on a single GPU?

Yes. At FP8 (E4M3) the model needs 30.4 GB of VRAM, so it fits on a single A100 40GB SXM4 or anything with more VRAM. At full FP16 it needs 60.4 GB, which fits on one Mac M3 Max (64GB) or larger.

How many GPUs does Qwen 3 30B-A3B (MoE) need?

At FP8 (E4M3) with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 4× RTX 3080 10GB (40 GB total VRAM). $0.005513 per 1M output tokens.

How fast can Qwen 3 30B-A3B (MoE) be served?

On 1× Cerebras CS-3 (WSE-3) at FP8 (E4M3), the aggregate throughput is 78000 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 Qwen 3 30B-A3B (MoE) need?

Weights alone: 60.0 GB at FP16, 30.0 GB at FP8 (E4M3). 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 Qwen 3 30B-A3B (MoE) in FP8 or FP16?

FP8 halves the weight footprint (60 GB → 30 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 Qwen 3 30B-A3B (MoE) need more VRAM than its active parameter count suggests?

Qwen 3 30B-A3B (MoE) is a mixture-of-experts model: 3B parameters are active per token, but all 30B 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 Qwen 3 30B-A3B (MoE) than to call an API?

At 4× RTX 3080 10GB the marginal cost is $0.005513 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 Qwen 3 30B-A3B (MoE) 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

Qwen 3 30B-A3B (MoE) on specific GPUs

Qwen 3 30B-A3B (MoE) on RTX 5090 32GBQwen 3 30B-A3B (MoE) on AMD MI300X 192GBQwen 3 30B-A3B (MoE) on A100 40GB SXM4Qwen 3 30B-A3B (MoE) on AMD MI325X 288GBQwen 3 30B-A3B (MoE) on A100 80GB SXM4Qwen 3 30B-A3B (MoE) on Intel Gaudi 3 128GBQwen 3 30B-A3B (MoE) on H100 SXM5 80GBQwen 3 30B-A3B (MoE) on B200 SXM 192GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters30.00B
Active parameters per token3.00B
ArchitectureMixture-of-Experts
Attention layoutFull attention (GQA)
Transformer layers48
Hidden dimension2048
Attention heads32 query / 4 KV (GQA)
Head dimension128
Vocabulary size151,936
Max context (catalog)41K
KV cache per token @ 8K96.0 KB

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