Skip to content

Llama Guard 4 12B on M3 Ultra 96 GB: VRAM, speed and fit

Llama Guard 4 12B at 8-bit runs on the M3 Ultra 96 GB with 288K of context at about 39 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
90.7 GB
of 96.0 GB
Context
288K
auto
Speed
39 tok/s
7.4 when full
First reply
2.0 s
2K prompt

Where the memory goes

Model
11.9 GB
Context
54.0 GB
Overhead
811 MB
OS
24.0 GB
Free
5.3 GB
Memory
90.7 GB of 96.0 GB
Context
288K tokens · auto
Speed
39 tok/s · 7.4 when full
First reply
2.0 s for a 2K prompt
Running cost
$3.55/mo at home · $122/mo on a cloud 4× RTX A4500

Validation: Validated

What Llama Guard 4 12B on M3 Ultra 96 GB was checked against
CheckTargetResultStatus
Fits in GPU memory72.0 GB usable66.7 GB neededPass
Headroom5% free5.3 GB free (7%)Pass
Context for chat16K or more288KPass
Writing speed15 tok/s or more39 tok/sPass
First replyunder 3 s, 2K prompt2.0 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
Llama Guard 4 12B
Parameters
12B
Precision
8-bit (Q8_0)
Context
288K, 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 22, 2026.

How to run it

llama.cpp
llama-server -hf DevQuasar/meta-llama.Llama-Guard-4-12B-GGUF
vLLM
vllm serve meta-llama/Llama-Guard-4-12B

Llama Guard 4 12B on similar machines

All machines: Llama Guard 4 12B 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 Llama Guard 4 12B run on M3 Ultra 96 GB?

Llama Guard 4 12B at 8-bit runs on the M3 Ultra 96 GB with 288K of context at about 39 tokens a second.

Report VR-0EP101W, calculated Nov 22, 2026 from published specs, not benchmarked.