Local AI on the MacBook: What Runs on 8 GB

Local AI on the Cheapest MacBook: What Runs on 8 GB

Short answer: Yes, the cheapest MacBook can run real local AI — if you pick the right version of each model. On an 8 GB MacBook Neo, Stable Diffusion 1.5 and SDXL Turbo worked straight away, FLUX and music generation worked with smaller configured versions, and the big default checkpoints did not fit. Anything that doesn’t fit can run in the Promptus online lane instead.

The Mac machine

A18 Pro chip — the same one as the iPhone 16 Pro family.
8 GB of unified memory, about 60 GB/s of memory bandwidth. There is no bigger Neo to buy.
Promptus runs ComfyUI locally (on port 8288) with Cosy alongside it. ComfyUI reported Apple’s Metal (MPS) path with the full 8 GB visible.
Internet is needed once, to download the app and the model files. After that, every local test ran offline.

What worked

Model	Setup	Time
Stable Diffusion 1.5	768×768, 20 steps	3 min 42 s
SDXL Turbo	7 GB checkpoint, 1 step, 512×512	4 min 45 s
FLUX (quantised)	Smaller quantised model, 256×256, decoder on the CPU	About 5 min
Music (smaller ACE-Step setup)	10 seconds of audio, both text encoders on the CPU	Just under 8 min

The pattern: the same 8 GB machine succeeded or failed depending on which version of a model was picked, not the model family. The quantised FLUX worked; the full-size one didn’t.

What didn’t fit

Model	What happened
FLUX Schnell, full 17 GB checkpoint	Out-of-memory (MPS) error after about 27 seconds
Z-Image, Q2 size (about 4 GB), 512×512, 4 steps	Still running after 10 minutes, no image, no error
FLUX.2 Klein	A data-format error on the Mac graphics backend at about 3.5 minutes; patched, but still no image after 10 minutes
Stock ACE music preset	Ran for 10 minutes without finishing

Those are the results as tested in this video. Z-Image Turbo was later run on an 8 GB Mac with a low-memory recipe — compressed models plus a memory helper node — covered in Z-Image Turbo for beginners on an 8 GB Mac.

Built-in tools worth using

Prompt Enhance turns a short idea into a detailed prompt before you generate.
Prompt history keeps every prompt one click away.
Image Describer reads a picture back to you — useful to check the result matches what you meant.

When it doesn’t fit: the online lane

Same account, same catalogue — switch to online. In the video, a muscle car through Seedream, a fighter jet through FLUX, and a 5-second video of the car through Seedance all finished in the cloud. Video in particular is something this machine was never going to render locally.

Automating it

Cosy exposes a local API on port 8190. The video listed the installed workflows, submitted a generation as a plain JSON request and polled it until it finished — enough to hand the job to a script or a coding agent.

Verdict

Start with Stable Diffusion 1.5, then SDXL Turbo. Move to the smaller FLUX and music setups once those work, and use the online lane for video and the biggest models. More on Mac options: Pinokio vs Promptus on a Mac. Download Promptus for Mac.

Frequently Asked Questions

Can a MacBook Neo with 8 GB run local AI image generation?

Yes, within limits. In the video, Stable Diffusion 1.5 and SDXL Turbo worked out of the box, and FLUX and music generation worked with smaller, specifically configured versions. Larger defaults did not fit in 8 GB.

How long does Stable Diffusion 1.5 take on a MacBook Neo?

A 768×768 image at 20 steps took 3 minutes 42 seconds, fully offline, on the A18 Pro with 8 GB of unified memory.

Which models failed on the 8 GB MacBook Neo?

The full 17 GB FLUX Schnell checkpoint hit an out-of-memory error in about 27 seconds. Z-Image at the Q2 size, FLUX.2 Klein and the stock ACE music preset ran for 10 minutes without finishing.

How do I run models that don’t fit on my Mac?

Switch Promptus to its online lane. In the video, Seedream, FLUX and a 5-second Seedance video all finished in the cloud from the same account and catalogue.

Can I automate Promptus on a Mac?

Yes. Cosy exposes a local API on port 8190. The video listed installed workflows, submitted a generation as a JSON request and polled it until it finished.

Written by:

Claudia Perez

Claudia Perez is a tech creator and founder at Promptus. She loves to design and share expressive image workflows. Her work focuses on visual storytelling, refined aesthetics, and ComfyUI workflows.

Try Promptus Cosy UI today for free.

Most recent wikis

One reference character rendered from eight camera angles in Promptus