Skip to content

Diffusiongemma 26B A4B IT on Mac mini M4 32 GB: VRAM, speed and fit

Diffusiongemma 26B A4B IT at 5-bit runs on the M4 32 GB with 104K of context at about 30 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on M4 32 GB · 5-bit (Q5_K_M) · Ollama / LM Studio · Chat

GPU memory
30.8 GB
of 32.0 GB
Context
104K
auto
Speed
30 tok/s
17 when full
First reply
16 s
2K prompt

Where the memory goes

Model
17.1 GB
Context
2.2 GB
Overhead
856 MB
OS
10.7 GB
Free
1.2 GB
Memory
30.8 GB of 32.0 GB
Context
104K tokens · auto
Speed
30 tok/s · 17 when full
First reply
16 s for a 2K prompt
Running cost
$0.85/mo at home · $33/mo on a cloud RTX A5000

Validation: Not validated

What Diffusiongemma 26B A4B IT on Mac mini M4 32 GB was checked against
CheckTargetResultStatus
Fits in GPU memory21.3 GB usable20.2 GB neededPass
Headroom5% free1.2 GB free (5%)Pass
Context for chat16K or more104KPass
Writing speed15 tok/s or more30 tok/sPass
First replyunder 3 s, 2K prompt16 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
Diffusiongemma 26B A4B IT
Parameters
25.8B, 4B active
Precision
5-bit (Q5_K_M)
Context
104K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
M4 32 GB
GPU can use
21.3 GB of 32.0 GB

Running cost

At home
$0.85/mo
39 W at the wall
Renting instead
$33/mo
RTX A5000 at $0.27/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 unsloth/diffusiongemma-26B-A4B-it-GGUF
vLLM
vllm serve google/diffusiongemma-26B-A4B-it
MLX (Mac)
mlx_lm.chat --model mlx-community/diffusiongemma-26B-A4B-it-4bit

Diffusiongemma 26B A4B IT on similar machines

Other models that run on Mac mini M4 32 GB

Questions

Can Diffusiongemma 26B A4B IT run on Mac mini M4 32 GB?

Diffusiongemma 26B A4B IT at 5-bit runs on the M4 32 GB with 104K of context at about 30 tokens a second.

Report VR-1UNEDQ5, calculated Nov 22, 2026 from published specs, not benchmarked.