Skip to content

Qwen2.5 VL 32B Instruct on 4× RTX 3090: VRAM, speed and fit

Qwen2.5 VL 32B Instruct at 8-bit runs on the 4× RTX 3090 with 125K of context at about 18 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on 4× RTX 3090 · 8-bit (Q8_0) · Ollama / LM Studio · Chat

GPU memory
67.3 GB
of 96.0 GB
Context
125K
auto
Speed
18 tok/s
9.4 when full
First reply
0.6 s
2K prompt

Where the memory goes

Model
33.1 GB
Context
31.3 GB
Overhead
2.9 GB
OS
0.0 GB
Free
28.7 GB
Memory
67.3 GB of 96.0 GB
Context
125K tokens · auto
Speed
18 tok/s · 9.4 when full
First reply
0.6 s for a 2K prompt
Running cost
$20/mo at home · $122/mo on a cloud 4× RTX A4500

Validation: Validated

What Qwen2.5 VL 32B Instruct on 4× RTX 3090 was checked against
CheckTargetResultStatus
Fits in GPU memory96.0 GB usable67.3 GB neededPass
Headroom5% free28.7 GB free (30%)Pass
Context for chat16K or more125KPass
Writing speed15 tok/s or more18 tok/sPass
First replyunder 3 s, 2K prompt0.6 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
Qwen2.5 VL 32B Instruct
Parameters
33.5B
Precision
8-bit (Q8_0)
Context
125K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
4× RTX 3090
GPU can use
96.0 GB of 96.0 GB

Running cost

At home
$20/mo
920 W at the wall
Renting instead
$122/mo
4× RTX A4500 at $1.00/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.

How to run it

llama.cpp
llama-server -hf bartowski/Qwen_Qwen2.5-VL-32B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen2.5-VL-32B-Instruct

Qwen2.5 VL 32B Instruct on similar machines

All machines: Qwen2.5 VL 32B Instruct VRAM requirements. Everything that runs here: what can the 4× RTX 3090 run?

Other models that run on 4× RTX 3090

Questions

Can Qwen2.5 VL 32B Instruct run on 4× RTX 3090?

Qwen2.5 VL 32B Instruct at 8-bit runs on the 4× RTX 3090 with 125K of context at about 18 tokens a second.

Report VR-0B2UZPC, calculated Nov 22, 2026 from published specs, not benchmarked.