LLMBOTTLENECK.COM
Qwen / qwen2

Qwen2.5-72B-Instruct

72.7 billion parameters.

Open in the calculator →

Architecturepublished data
Quick answer
Qwen2.5-72B-Instruct needs about 47.3 GB at Q4_K_M — the format most people download — with 8,192 tokens of context. It fits entirely on 8 of 18 common devices, starting with the NVIDIA RTX 6000 Ada Generation (48 GB). What else fits in 48 GB. No hardware for it? Rent a GPU that runs it, priced live.

Sized at 8,192 tokens of context with the whole model in device memory, across 18 common devices. Speeds are estimates, not benchmarks.

Find the NVIDIA RTX 6000 Ada Generation, the smallest card listed here that holds it: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Weight-file size by format

FormatSizeRangeBasis
FP16145.4 GB142.9 GB – 146.1 GBreconstructed size
Q8_077.3 GB75.9 GB – 77.6 GBreconstructed size
Q6_K59.7 GB58.6 GB – 59.9 GBreconstructed size
Q5_K_M51.5 GB50.5 GB – 51.7 GBreconstructed size
Q5_050.2 GB49.1 GB – 50.4 GBreconstructed size
Q4_K_M43.8 GB42.8 GB – 44.1 GBreconstructed size
Q4_041.2 GB40.2 GB – 41.5 GBreconstructed size
Q3_K_M35.5 GBexactpublished data
Q2_K27.3 GBexactpublished data

2 of these are published files. Open the pinned GGUF repository ↗

A file size is not the memory a run needs: the KV cache and the runtime reserve come on top, and the calculator adds both.

What the labels mean
published data
Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
reconstructed size
Weight size reconstructed from the pinned architecture, because no published file exists.
size range
Only a lower and an upper bound are claimed for this weight size.

Hardware ladder

Q4_K_M at 8,192 tokens · 8 of 18 devices hold it entirely
DeviceMemoryVerdictDecodeCalculator
NVIDIA GeForce RTX 507047.3 GB of 12 GB65 layers on system RAM~1.6 tok/s with offloadOpen →
AMD Radeon™ RX 9070 XT47.3 GB of 16 GB58 layers on system RAM~1.8 tok/s with offloadOpen →
NVIDIA GeForce RTX 408047.3 GB of 16 GB58 layers on system RAM~1.8 tok/s with offloadOpen →
NVIDIA GeForce RTX 508047.3 GB of 16 GB58 layers on system RAM~1.8 tok/s with offloadOpen →
NVIDIA GeForce RTX 5070 Ti47.3 GB of 16 GB58 layers on system RAM~1.8 tok/s with offloadOpen →
NVIDIA GeForce RTX 5060 Ti 16GB47.3 GB of 16 GB58 layers on system RAM~1.8 tok/s with offloadOpen →
NVIDIA GeForce RTX 409047.3 GB of 24 GB43 layers on system RAM~2.4 tok/s with offloadOpen →
AMD Radeon™ RX 7900 XTX47.3 GB of 24 GB43 layers on system RAM~2.4 tok/s with offloadOpen →
NVIDIA GeForce RTX 309047.3 GB of 24 GB43 layers on system RAM~2.4 tok/s with offloadOpen →
NVIDIA GeForce RTX 509047.3 GB of 32 GB28 layers on system RAM~3.6 tok/s with offloadOpen →
NVIDIA RTX 6000 Ada Generation47.3 GB of 48 GBfits~14 tok/sAbout reading paceOpen →
Apple M4 Pro47.3 GB of 64 GBfits~5.4 tok/sSlowOpen →
Apple M5 Pro47.3 GB of 64 GBfits~6.1 tok/sSlowOpen →
AMD Ryzen AI Max+ 395 with Radeon 8060S47.3 GB of 128 GBfits~4.4 tok/sSlowOpen →
Apple M3 Max47.3 GB of 128 GBfits~7.8 tok/sSlowOpen →
Apple M4 Max47.3 GB of 128 GBfits~10 tok/sAbout reading paceOpen →
NVIDIA DGX Spark47.3 GB of 128 GBfits~4.0 tok/sSlowOpen →
Apple M2 Ultra47.3 GB of 192 GBfits~15 tok/sAbout reading paceOpen →

Decode is a calibrated estimate from the published calibration; capacity is the manufacturer's published ceiling, not guaranteed free memory. Fitting in memory is not the same as loading: whether the runtime and version you have supports this architecture and format on that machine has not been tested here. Offloaded rows assume there is enough system RAM for the overflow — the calculator checks that against the RAM you declare. Each “Open” link keeps this model, format and context.

Run it in the cloud

Rent a GPU that runs Qwen2.5-72B-Instruct

Every rentable machine that holds the whole model, from Vast.ai and RunPod, with this site's speed estimate for each and the provider's own price, read live. Cheapest first; sort by cost per token or speed instead, or switch the format.

Prices · 18:30 UTC, 22 Sept
19 configurations in stock · 14 out of stock
MachineSpeedPer hourPer M tokensWhereRent
1× RTX 6000 Ada48 GBcheapest~14 tok/sAbout reading pace$0.26$5.10Vast.aiMarketplace · 99.9% reliableRunPod $0.84/h Rent on Vast.ai
1× RTX A600048 GB~11 tok/sAbout reading pace$0.28$6.86Vast.aiMarketplace · 99.6% reliableRunPod $0.53/h Rent on Vast.ai
1× RTX PRO 500048 GB~19 tok/sAbout reading pace$0.51$7.10Vast.aiMarketplace · 98.1% reliableRunPod $0.82/h Rent on Vast.ai
4× RTX 5060 Ti 16GB64 GB~26 tok/sAbout reading pace$0.51$5.31Vast.aiMarketplace · 99.5% reliableRent on Vast.ai
2× RTX 509064 GBrecommendedbest value~53 tok/sFaster than you read$0.54$2.80Vast.aiMarketplace · 99.6% reliableRent on Vast.ai
8× RTX 306096 GB~42 tok/sFaster than you read$0.59$3.83Vast.aiMarketplace · 99.7% reliableRent on Vast.ai
4× RTX 309096 GB~55 tok/sFaster than you read$0.60$2.98Vast.aiMarketplace · 99.4% reliableRent on Vast.ai
1× L40S48 GB~12 tok/sAbout reading pace$0.69$15.08Vast.aiMarketplace · 98.5% reliableRunPod $0.79/h Rent on Vast.ai

No GPU at all

Use Qwen2.5-72B-Instruct by the token

For one person chatting, that is about 7.0× cheaper than the best-value rented card above ($2.80 per million tokens). A rented GPU pays off when you keep it busy — many requests batched together, long agent runs — or when the data must stay on a machine you control, or the exact file you want is not served anywhere.

First time renting a GPU? How it works, in four steps
  1. Create an account and add credit. Both providers are prepaid: $10 is enough to try any card on this page for hours. A new RunPod account that arrives through a referral link gets a one-time $5 credit when it first adds $10.
  2. Pick the card and the count shown here, and a template: llama.cpp or vLLM for an OpenAI-compatible API, or one with a web chat if you only want to talk to the model. “How to launch it” below gives the exact image and command.
  3. Wait for the download. The weights are fetched on the machine; tens of gigabytes take a few minutes on a data-centre connection. You pay by the second from the moment it starts.
  4. Stop it when you are done. A running machine bills even when idle. A stopped one costs nothing per hour, though its disk is usually still billed until you delete it.
How to launch it

No Q4_K_M file of Qwen2.5-72B-Instruct is catalogued here, so there is no exact command to vouch for. Start the machine from the llama.cpp server image ghcr.io/ggml-org/llama.cpp:server-cuda and point -hf at a Q4_K_M build from the publisher or a quantizer you trust — search Hugging Face for one. Check its file size against the memory figure on this page before you rent.

Every machine holds the whole model at Q4_K_M and 8,192 tokens of context, no offload. Speeds are this site’s single-stream decode estimates; prices are what each provider’s own API quoted, on-demand, for the whole machine. Vast hosts below 98% measured reliability are left out.

Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank

Buying a card instead, or paying by the token? Compare a year of Qwen2.5-72B-Instruct three ways →

Under the hood

Architecture, read from the publisher’s file

The numbers every figure above is computed from, with the file they came from.

Architecture

✓ Architecture read from the published config.json

Retrieved 2026-09-01 at pinned commit 495f39366efe; parameter count from the safetensors index at the same revision.

Architecture
qwen2
Layers
80
Hidden size
8,192
Attention heads
64
KV heads
8
Feed-forward width
29,568
Vocabulary
152,064
Context ceiling
32,768
Sliding window
131,072
RoPE theta
1,000,000

Where the memory goes

tokenembedding× 80 decoder blocksGrouped-query attention64 query · 8 KV headswindow 131,072Feed-forwardone networkall activeoutputprojectiongrows with contextfixed per token

Grouped-query attention shares each key/value head across 8 query heads, so the KV cache is 13% of what multi-head attention would need at the same context.

Open the exact official config.json ↗

More from Qwen

Qwen-AgentWorld-35B-A3BQwen2.5-0.5BQwen2.5-0.5B-InstructQwen2.5-1.5B-InstructQwen2.5-14B-InstructQwen2.5-32B-InstructQwen2.5-3B-InstructQwen2.5-7B-Instruct

Use this answer in your own product

As JSON, from the API
curl -s https://llmbottleneck.com/v1/analyze \
  -H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
  -d '{"model":"qwen-qwen2-5-72b-instruct","quantization":"Q4_K_M","context":8192,"hardware":"nvidia-geforce-rtx-4090"}'

Same engine, same evidence, every field sourced. Free key in one step →

On your page, as a widget
<script src="https://llmbottleneck.com/widget.js"
  data-model="qwen-qwen2-5-72b-instruct" data-quantization="Q4_K_M"
  data-hardware="nvidia-geforce-rtx-4090" data-context="8192"></script>

No key needed. Unbranded, with your own buy button, on Pro and Business →

Citing this page

LLM Bottleneck. “Qwen2.5-72B-Instruct VRAM and hardware requirements.” Architecture from Qwen/Qwen2.5-72B-Instruct at revision 495f39366efe, retrieved 2026-09-01. https://llmbottleneck.com/models/qwen-qwen2-5-72b-instruct

Every figure above is either the published value or a reconstruction whose measured error is on the accuracy page.

llmbottleneck
catalogue 2026-10-03models 327devices 135