Skip to content

Gemma 4 31B on M3 Ultra 96 GB: VRAM, speed and fit

Gemma 4 31B at 8-bit runs on the M3 Ultra 96 GB with 256K of context at about 15 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
66.6 GB
of 96.0 GB
Context
256K
auto
Speed
15 tok/s
11 when full
First reply
5.9 s
2K prompt

Where the memory goes

Model
30.9 GB
Context
10.8 GB
Overhead
936 MB
OS
24.0 GB
Free
29.4 GB
Memory
66.6 GB of 96.0 GB
Context
256K tokens · auto
Speed
15 tok/s · 11 when full
First reply
5.9 s for a 2K prompt
Running cost
$3.55/mo at home · $66/mo on a cloud 2× RTX A5000

Validation: Validated with conditions

What Gemma 4 31B on M3 Ultra 96 GB was checked against
CheckTargetResultStatus
Fits in GPU memory72.0 GB usable42.6 GB neededPass
Headroom5% free29.4 GB free (41%)Pass
Context for chat16K or more256KPass
Writing speed15 tok/s or more15 tok/sCheck
First replyunder 3 s, 2K prompt5.9 sCheck

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
Gemma 4 31B
Parameters
31.3B
Precision
8-bit (Q8_0)
Context
256K, 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
$66/mo
2× RTX A5000 at $0.54/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

Ollama
ollama run gemma4:31b
llama.cpp
llama-server -hf unsloth/gemma-4-31B-it-GGUF
vLLM
vllm serve google/gemma-4-31B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/gemma-4-31b-it-4bit

Gemma 4 31B on similar machines

All machines: Gemma 4 31B 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 Gemma 4 31B run on M3 Ultra 96 GB?

Gemma 4 31B at 8-bit runs on the M3 Ultra 96 GB with 256K of context at about 15 tokens a second.

Report VR-0Z7Z7RH, calculated Nov 22, 2026 from published specs, not benchmarked.