Skip to content

Qwen3 VL 32B Instruct on M3 Ultra 96 GB: VRAM, speed and fit

Qwen3 VL 32B Instruct at 8-bit runs on the M3 Ultra 96 GB with 136K of context at about 14 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on M3 Ultra 96 GB · 8-bit (Q8_0) · Ollama / LM Studio · Chat

GPU memory
91.7 GB
of 96.0 GB
Context
136K
auto
Speed
14 tok/s
7.2 when full
First reply
6.2 s
2K prompt

Where the memory goes

Model
33.0 GB
Context
34.0 GB
Overhead
713 MB
OS
24.0 GB
Free
4.3 GB
Memory
91.7 GB of 96.0 GB
Context
136K tokens · auto
Speed
14 tok/s · 7.2 when full
First reply
6.2 s for a 2K prompt
Running cost
$3.55/mo at home · $122/mo on a cloud 4× RTX A4500

Validation: Not validated

What Qwen3 VL 32B Instruct on M3 Ultra 96 GB was checked against
CheckTargetResultStatus
Fits in GPU memory72.0 GB usable67.7 GB neededPass
Headroom5% free4.3 GB free (6%)Pass
Context for chat16K or more136KPass
Writing speed15 tok/s or more14 tok/sCheck
First replyunder 3 s, 2K prompt6.2 sFails

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
Qwen3 VL 32B Instruct
Parameters
33.4B
Precision
8-bit (Q8_0)
Context
136K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
M3 Ultra 96 GB
GPU can use
72.0 GB of 96.0 GB

Running cost

At home
$3.55/mo
162 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 Qwen/Qwen3-VL-32B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen3-VL-32B-Instruct

Qwen3 VL 32B Instruct on similar machines

All machines: Qwen3 VL 32B Instruct VRAM requirements. Everything that runs here: what can the M3 Ultra 96 GB run?

Other models that run on M3 Ultra 96 GB

Questions

Can Qwen3 VL 32B Instruct run on M3 Ultra 96 GB?

Qwen3 VL 32B Instruct at 8-bit runs on the M3 Ultra 96 GB with 136K of context at about 14 tokens a second.

Report VR-0PFDZ3C, calculated Nov 22, 2026 from published specs, not benchmarked.