Vision-languageDense72B params128K max context

Qwen 2.5-VL 72B (VLM) VRAM Requirements: How Many GPUs & What It Costs

Qwen 2.5-VL 72B (VLM) needs 72.0 GB of VRAM for its weights at FP8 (E4M3) — 73.3 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))
72.0 GB
73.3 GB with KV cache
Fits on 1 GPU?
Yes
TPU v5p (per-chip, 95GB) or larger
Cheapest config
$0.1052
8× RTX 4080 16GB / 1M tokens
Fastest config
44200
tok/s on 4× 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
FP16144.0 GB145.3 GBB200 SXM 192GB (192 GB)
FP8 (E4M3)recommended72.0 GB73.3 GBH100 SXM5 80GB (80 GB)
AWQ 4-bit39.6 GB40.9 GBL40S 48GB (48 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 4080 16GB8×128 GB66.026338$0.1052
RTX 5090 32GB4×128 GB82.513960$0.1403
RTX 3090 24GB4×96 GB43.091978$0.1404
RTX A4000 16GB8×128 GB41.253960$0.1683
RTX 4090 24GB4×96 GB46.412228$0.1995
T4 16GB8×128 GB29.471800$0.2469
V100 SXM2 32GB4×128 GB41.441989$0.2793
A100 40GB SXM44×160 GB71.363426$0.2919
A10G 24GB4×96 GB27.631326$0.3017
TPU v6e Trillium (32GB)4×128 GB37.711810$0.4911
A100 80GB SXM42×160 GB46.041105$0.6033
L4 24GB4×96 GB13.81663.0$0.6704

RTX 4080 16GB 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 4080 16GB 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?
4K411.34 GB✓ fits
8K202.68 GB✓ fits
33K510.74 GB✓ fits
66K221.47 GB✓ fits
131K142.95 GB✓ fits
Frequently asked questions

Can Qwen 2.5-VL 72B (VLM) run on a single GPU?

Yes. At FP8 (E4M3) the model needs 73.3 GB of VRAM, so it fits on a single TPU v5p (per-chip, 95GB) or anything with more VRAM. At full FP16 it needs 145.3 GB, which fits on one B200 SXM 192GB or larger.

How many GPUs does Qwen 2.5-VL 72B (VLM) need?

At FP8 (E4M3) with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 4080 16GB (128 GB total VRAM). $0.1052 per 1M output tokens.

How fast can Qwen 2.5-VL 72B (VLM) be served?

On 4× Cerebras CS-3 (WSE-3) at FP8 (E4M3), the aggregate throughput is 44200 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 2.5-VL 72B (VLM) need?

Weights alone: 144.0 GB at FP16, 72.0 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 2.5-VL 72B (VLM) in FP8 or FP16?

FP8 halves the weight footprint (144 GB → 72 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.

Is it cheaper to self-host Qwen 2.5-VL 72B (VLM) than to call an API?

At 8× RTX 4080 16GB the marginal cost is $0.1052 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 2.5-VL 72B (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

Qwen 2.5-VL 72B (VLM) on specific GPUs

Qwen 2.5-VL 72B (VLM) on RTX 5090 32GBQwen 2.5-VL 72B (VLM) on A100 80GB SXM4Qwen 2.5-VL 72B (VLM) on H100 SXM5 80GBQwen 2.5-VL 72B (VLM) on AMD MI300X 192GBQwen 2.5-VL 72B (VLM) on L40S 48GBQwen 2.5-VL 72B (VLM) on AMD MI325X 288GBQwen 2.5-VL 72B (VLM) on Intel Gaudi 3 128GBQwen 2.5-VL 72B (VLM) on B200 SXM 192GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters72.00B
Active parameters per token72.00B
ArchitectureDense
Attention layoutFull attention (GQA)
Transformer layers80
Hidden dimension8192
Attention heads64 query / 8 KV (GQA)
Head dimension128
Vocabulary size152,064
Max context (catalog)128K
KV cache per token @ 8K320.0 KB

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