AlibabaQwen 2.5Qwen2ForCausalLM

Qwen2.5 72B — Hardware Requirements & GPU Compatibility

ChatFunctions

Qwen2.5-72B is the base, pretrained 72.7-billion-parameter language model from Alibaba's Qwen2.5 series — it is not instruction-tuned and is not intended for direct conversational use; Alibaba recommends applying SFT, RLHF, or further pretraining on top of it. It uses a Transformer architecture with RoPE, SwiGLU, RMSNorm, QKV attention bias, 80 layers, and grouped-query attention (64 query heads, 8 key/value heads), with stronger coding, math, and structured-output ability than the earlier Qwen2 line, plus multilingual coverage across 29-plus languages. At 72.7 billion parameters, it needs a multi-GPU workstation to run at full precision, though quantized versions fit fewer cards. Context length is 131,072 tokens. It is released under Alibaba's custom Qwen license, which permits research and commercial use but requires a separate license from Alibaba once a product or service passes 100 million monthly active users. It was published in September 2024.

72.7B params · dense131K contextOtherSep 20247.8K downloads114 likes6.4K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Doesn't fit a 24 GB GPU

Smallest consumer fit → Snapdragon X Elite Copilot+ PC · IQ2_XXS · 25.7 GB · ~4 tok/s

Macs from 32 GB (IQ2_S · ~12 tok/s)

Other

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Qwen2.5 72B

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended

Ollama

one command · needs Ollama installed

$ollama run hf.co/mradermacher/Qwen2.5-72B-i1-GGUF:Q4_K_M

Q4_K_M · needs ~49.3 GB · 16K context

Download: Qwen2.5-72B.i1-Q4_K_M.gguf47.42 GBmradermacher

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Your own private copy · Q4_K_M · 96 GB GPU

The easiest way to try this model
Running in a few minutes, easy setup
Chat right here or connect any app

$5+ bonus for new users

How Much VRAM Does Qwen2.5 72B Need?

Select a quantization level; the listed hardware adheres to it.

Quantization	VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
	21.0 GB	31 GB	63.3 GB	25.5 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 8K context · ~31 tok/s
	22.8 GB	32.9 GB	65.1 GB	27.1 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 4K context · ~29 tok/s
	23.7 GB	33.8 GB	66 GB	27.9 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 2K context · ~28 tok/s
	25.5 GB	35.6 GB	67.8 GB	29.3 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Does not fit: needs 30.2 GB
	29.1 GB	39.2 GB	71.4 GB	31.8 GB	3-bit · i-quant — Importance-weighted 3-bit	Does not fit: needs 33.8 GB
	30.1 GB	40.1 GB	72.3 GB	29.6 GB	2-bit · small K-quant — 2-bit small K-quant	Does not fit: needs 34.8 GB
	31.0 GB	41 GB	73.2 GB	32.8 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Does not fit: needs 35.7 GB
	31.9 GB	41.9 GB	74.2 GB	34.5 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Does not fit: needs 36.6 GB
	31.9 GB	41.9 GB	74.2 GB	29.8 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Does not fit: needs 36.6 GB
	32.8 GB	42.9 GB	75.1 GB	34.5 GB	3-bit small — 3-bit small quantization	Does not fit: needs 37.5 GB
	33.7 GB	43.8 GB	76 GB	35.5 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Does not fit: needs 38.4 GB
	36.4 GB	46.5 GB	78.7 GB	37.7 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 41.1 GB
	37.3 GB	47.4 GB	79.6 GB	41.2 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 42 GB
	38.2 GB	48.3 GB	80.5 GB	39.5 GB	3-bit large — 3-bit large quantization	Does not fit: needs 42.9 GB
	40.0 GB	50.1 GB	82.3 GB	39.7 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Does not fit: needs 44.8 GB
	41.9 GB	51.9 GB	84.2 GB	45.7 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Does not fit: needs 46.6 GB
	41.9 GB	51.9 GB	84.2 GB	43.9 GB	4-bit small — 4-bit small quantization	Does not fit: needs 46.6 GB
	41.9 GB	51.9 GB	84.2 GB	41.3 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Does not fit: needs 46.6 GB
	44.6 GB	54.7 GB	86.9 GB	47.4 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 49.3 GB
	51.0 GB	61 GB	93.2 GB	51.4 GB	5-bit small — 5-bit small quantization	Does not fit: needs 55.7 GB
	52.8 GB	62.8 GB	95.1 GB	54.4 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 57.5 GB
	61.0 GB	71 GB	103.2 GB	64.3 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 65.7 GB
	73.7 GB	83.8 GB	116 GB	77.3 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 78.4 GB
	146.4 GB	156.5 GB	188.7 GB	~145.4 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 151.1 GB

est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

49.3GBat 16K context

Weights: 43.6 GB
KV cache: 5.4 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 25.3 GB more

≥10 tok/s needs ≥683 GB/s · M3 Ultra ~13, M2 Ultra ~13

Or .

Which Hardware Can Run Qwen2.5 72B?

Q4_K_M · 44.6 GB

Qwen2.5 72B (Q4_K_M) requires 44.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 58+ GB is recommended. Using the full 131K context window can add up to 42.3 GB, bringing total usage to 86.9 GB. No consumer GPU has enough memory. 52 devices with unified memory can run Qwen2.5 72B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 25.7 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceDoesn't fit — smallest quant needs 25.7 GB24 GB tier →

Mac 32 GB · Mac Mini M4IQ2_S · 28.4 GB · ~4 tok/s · 16K28.4~416KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 49.3 GB · ~9 tok/s · 32K49.3~932KMacs →

How good is Qwen2.5 72B?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

MMLU
Epoch AI · Sep 2024Lab-reported
85.0%
#4/85 open
Lab-reported
BIG-Bench Hard
Epoch AI · Sep 2024Lab-reported
79.8%
#4/38 open
Lab-reported
HellaSwag
Epoch AI · Sep 2024Lab-reported
84.8%
#8/47 open
Lab-reported
ARC Challenge
Epoch AI · Sep 2024Lab-reported
94.5%
#3/56 open
Lab-reported
WinoGrande
Epoch AI · Sep 2024Lab-reported
82.3%
#6/51 open
Lab-reported
TriviaQA
Epoch AI · Sep 2024Lab-reported
71.9%
#12/19 open
Lab-reported

All 7 Qwen2.5 72B benchmarks

Alternatives to Qwen2.5 72B

Qwen 2.5 family · 44 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Llama 3.3 70B InstructMeta	80.0	51.3 GB	—needs > 24 GB	same size
NVIDIA Nemotron 3.5 Bolt 74B A9B BF16NVIDIA	—	54.2 GB	—needs > 24 GB	same size
Kolibri 1 BF16Aleph-Alpha	—	22.5 GB	~237	same size
CodeLlama 70B Instruct HFMeta	—	23.2 GB	~29	same size
StableBeluga2Stability AI	—	23.5 GB	~29	same size
Apertus 70B 2509swiss-ai	—	48 GB	—needs > 24 GB	same size

Fine-tunes & variants(2)community builds of Qwen2.5 72B

Specifications

Publisher: Alibaba
Family: Qwen 2.5
Parameters: 72.7B
Architecture: Qwen2ForCausalLM
Context Length: 131,072 tokens
Vocabulary Size: 152,064
Release Date: 2024-09-15
License: Other

Resources

Frequently Asked Questions

How much VRAM does Qwen2.5 72B need?

Qwen2.5 72B requires 44.6 GB of VRAM at Q4_K_M, or 146.4 GB at BF16. Full 131K context adds up to 42.3 GB (86.9 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 72.7B × 4.8 bits ÷ 8 = 43.6 GB

KV Cache + Overhead ≈ 1 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 43.3 GB (at full 131K context)

VRAM usage by quantization

Q4_K_M

44.6 GB

Q4_K_M + full context

86.9 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Qwen2.5 72B?

Yes, at IQ2_S (23.7 GB) or lower. Higher quantizations like IQ2_M (25.5 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Qwen2.5 72B?

For Qwen2.5 72B, Q4_K_M (44.6 GB) offers the best balance of quality and VRAM usage. Q5_K_S (51.0 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 21.0 GB.

VRAM requirement by quantization

IQ2_XXS

21.0 GB~53%

IQ3_XS

31.0 GB~73%

Q4_0

37.3 GB~85%

IQ4_NL

41.9 GB~88%

Q4_K_M ★

44.6 GB~89%

BF16

146.4 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Qwen2.5 72B on a Mac?

Yes — Mac Studio M1 Max (32 GB) and 84 other Macs can run Qwen2.5 72B, the smallest at IQ2_S. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 44.6 GB of usable unified memory (RAM minus macOS overhead), from Mac Pro M2 Ultra (64 GB).

Can I run Qwen2.5 72B locally?

Yes — Qwen2.5 72B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Qwen2.5 72B?

At Q4_K_M, Qwen2.5 72B can reach ~108 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 44.6 × 0.65 = ~117 tok/s

Estimated speed at Q4_K_M (44.6 GB)

NVIDIA B200

~117 tok/s

NVIDIA B300

~117 tok/s

AMD Instinct MI350X

~108 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Qwen2.5 72B?

At Q4_K_M, the download is about 43.62 GB. The full-precision BF16 version is 145.41 GB. The smallest option (IQ2_XXS) is 19.99 GB.

Which GPUs can run Qwen2.5 72B?

No single consumer GPU has enough VRAM to run Qwen2.5 72B at Q4_K_M (44.6 GB). Multi-GPU or professional hardware is required.

Which devices can run Qwen2.5 72B?

52 devices with unified memory can run Qwen2.5 72B at Q4_K_M (44.6 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.

Run Qwen2.5 72B

49.3 GB VRAM at 16K context

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/mradermacher/Qwen2.5-72B-i1-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/mradermacher/Qwen2.5-72B-i1-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/mradermacher/Qwen2.5-72B-i1-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf mradermacher/Qwen2.5-72B-i1-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/mradermacher/Qwen2.5-72B-i1-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf mradermacher/Qwen2.5-72B-i1-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve Qwen/Qwen2.5-72B --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~160 GB of GPU memory.

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$mradermacher/Qwen2.5-72B-i1-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
A100 80GBCheapestCheapest30 GB free	~30 tok/s	$10.62	Deploy on Vast.aifrom $1.14/hr (opens in new tab)	30 GB free · ~30 tok/s · $10.62 / 1M tokens
RTX PRO 6000 Max-Q (96 GB)46 GB free	~26 tok/s	$14.17	Deploy on Vast.aifrom $1.33/hr (opens in new tab)	96 GB · 46 GB free · ~26 tok/s · $14.17 / 1M tokens
A100 80GB30 GB free	~30 tok/s	$13.00	Deploy on RunPodfrom $1.39/hr (opens in new tab)	30 GB free · ~30 tok/s · $13.00 / 1M tokens
RTX PRO 6000 (96 GB)46 GB free	~26 tok/s	$16.34	Deploy on Vast.aifrom $1.54/hr (opens in new tab)	96 GB · 46 GB free · ~26 tok/s · $16.34 / 1M tokens
RTX PRO 6000 Max-Q (96 GB)46 GB free	~26 tok/s	$17.45	Deploy on RunPodfrom $1.64/hr (opens in new tab)	96 GB · 46 GB free · ~26 tok/s · $17.45 / 1M tokens
RTX PRO 6000 (96 GB)46 GB free	~26 tok/s	$17.99	Deploy on RunPodfrom $1.69/hr (opens in new tab)	96 GB · 46 GB free · ~26 tok/s · $17.99 / 1M tokens
RTX PRO 6000 Server (96 GB)46 GB free	~23 tok/s	$20.15	Deploy on RunPodfrom $1.69/hr (opens in new tab)	96 GB · 46 GB free · ~23 tok/s · $20.15 / 1M tokens
RTX PRO 6000 Server (96 GB)46 GB free	~23 tok/s	$20.72	Deploy on Vast.aifrom $1.74/hr (opens in new tab)	96 GB · 46 GB free · ~23 tok/s · $20.72 / 1M tokens

Compare GPUs →

Your own private copy of Qwen2.5 72B · Q4_K_M · OpenAI-compatible API

The pod’s first start downloads ~44 GB, billed while it loads.

Open the pod on RunPod
Deploy on RunPodRTX PRO 6000 · $2.49/hr (opens in new tab)
New to RunPod? Sign up and add credit first. If that GPU is gone, pick another with at least 96 GB.
2
Set the model and your key
Press Set overrides → Environment variables. The first variable is already there; set its value:
Key
LLAMA_ARG_HF_REPO
Value
mradermacher/Qwen2.5-72B-GGUF:Q4_K_M
Press Add Environment variable for your key:
Key
LLAMA_API_KEY
Value
Make up your own key
A long password works. You’ll paste it below as the API key. Without it, anyone who finds the pod’s address can use it.Then Set overrides.
Deploy
Press Deploy Pod. Usually ready in 15–60 min.

Use it

RunPod → your pod → Connect → HTTP 8080, copy the link. The key is the one you made up in step 2.

Endpoint URL

API key

Chats and keys stay only in this browser.

https://<POD_ID>-8080.proxy.runpod.net/v1

RunPod → your pod → Connect → HTTP 8080 gives this address. The API key is the one from step 2.

from openai import OpenAI
client = OpenAI(base_url="https://<POD_ID>-8080.proxy.runpod.net/v1", api_key="YOUR_KEY")
r = client.chat.completions.create(
    model="any",
    messages=[{"role": "user", "content": "Hello!"}])
print(r.choices[0].message.content)

Works with any OpenAI client.

https://<POD_ID>-8080.proxy.runpod.net/v1

RunPod → your pod → Connect → HTTP 8080 gives this address. The API key is the one from step 2.

Free on Windows, macOS, Linux, iOS, Android and the web.

Open the web app ↗ (opens in a new tab)Download ↗ (opens in a new tab)

1. Settings → Model Provider → Add → Add Custom Provider
2. API Mode OpenAI API Compatible; paste the API key and the base URL as API Host
3. Model → Fetch, pick the model
4. New chat → pick the provider → start chatting

Free and open source for Windows, macOS and Linux.

Download Jan ↗ (opens in a new tab)

1.Settings → next to Model Providers press +
2.API format OpenAI-compatible; Base URL is the base URL above (with /v1); API key is your key; Create
3.If no model appears, click + in Models and type any
4.New chat → pick the model → start chatting