How to Run MOSS-Image on 8GB

MOSS Image AI Edit on ComfyUI Workflow

Short answer: Yes, MOSS-Image runs on a computer with 8 GB of memory. It uses a 4-bit copy of the image model and a 3-bit copy of the text encoder, run through stable-diffusion.cpp. In Promptus, the ready-made MOSS-Image 8GB Fast CosyTemplate does the setup for you. On an 8 GB MacBook (Apple A18 Pro), the fast 4-step version made a 512×512 picture in 2 minutes 16 seconds, with the sign spelled correctly. The full model took 10 minutes 44 seconds.

Watch: MOSS-Image on 8 GB — the full walkthrough on YouTube · Get it: open the MOSS-Image 8GB Fast CosyTemplate, or in Promptus go to CosyTemplates › search “MOSS” › Install Locally · Download Promptus

Four pictures made with MOSS-Image on an 8 GB MacBook: a chalkboard reading PROMPTUS, a JUNO LIVE gig poster, a WILD HONEY jar label and an OPEN neon sign in English and Chinese — All four made on an 8 GB MacBook with the MOSS-Image 8GB CosyTemplate. The big printed words are right; the small handwriting on the chalkboard is not — see the tips below.

What is MOSS-Image?

MOSS-Image is an open-source text-to-image model from the OpenMOSS team, released on 16 October 2026 alongside its technical report. It was built for the thing most image models get wrong: text inside the picture. Signs, posters, menus and labels come out spelled correctly, in English and Chinese.

	MOSS-Image
Made by	OpenMOSS
Released	16 October 2026
Size	5 billion parameters
Licence	Apache-2.0 (commercial use allowed)
Prompt length	Up to 4,096 tokens
Special features	Text in English and Chinese; layout control with JSON boxes

Why is MOSS-Image so popular?

On 26 October 2026 it was the #1 trending text-to-image model on Hugging Face, and it was still second on 27 October.
The ComfyUI version on Hugging Face showed 2,760,550 downloads on 27 October, eleven days after it appeared. Hugging Face counts file downloads, not people.
On LongLongTextBench, a long-text test made by the MOSS team, it scored 96.3% average text coverage, the best of the open models they tested. Coverage measures how much of the requested text appears, not whether every letter is right.

Why the normal workflow needs about 13 GB, and how the 8GB version fits

The official ComfyUI files are a 5.4 GB image model and a 7.3 GB text encoder (the part that reads your prompt). That’s about 13 GB, more than an 8 GB computer has once the system is running. The 8GB version swaps in smaller copies of the same model:

Part	Official ComfyUI file	8GB version
Image model	int8, 5.4 GB	GGUF Q4_0, 3.3 GB (the 4-step version is the same size)
Text encoder (Qwen3.5-9B)	w4a8, 7.3 GB	GGUF Q3_K_S, 4.3 GB
VAE (FLUX.2)	0.34 GB	Same file
Engine	ComfyUI	stable-diffusion.cpp, official release build

Three things keep memory down:

The files are memory-mapped, so only the part in use sits in memory.
The text encoder runs on the processor, which leaves the graphics memory for the image model.
The picture is decoded in tiles.

Nothing has to be compiled. The first time you press Run, the node downloads the ready-made engine for your computer (about 30 MB) and the model files (about 7.9 GB).

How to run MOSS-Image in Promptus on 8GB

CosyTemplates:

MOSS-Image 8GB Fast Text to Image — 4 steps, about 2 minutes per 512×512 picture on 8 GB. Start here.
MOSS-Image 8GB Text to Image — the full model, 16 steps, about 11 minutes per 512×512 picture on 8 GB.

Install it and make a picture (no GitHub, no Terminal)

In Promptus, open CosyTemplates and search for MOSS.
On the MOSS-Image 8GB Fast Text to Image card, open the menu and choose Install Locally. Promptus copies the workflow to your computer and fetches its listed models.
When it shows under Available Locally, open the menu again and choose Load in ComfyUI.
Type your prompt. Put the exact words you want written in “quotation marks”.
Press Run. The first run downloads the engine and the rest of the model files (about 8 GB in total), so it takes longer. After that, it’s about 2 minutes a picture on 8 GB.
If ComfyUI says the MOSS-Image node is missing, click Install in that message, then restart ComfyUI.

MOSS-Image 8GB workflow loaded in ComfyUI from the CosyTemplate — Step 3: the workflow after Load in ComfyUI

Using ComfyUI without Promptus? Install the node with ComfyUI Manager › Install via Git URL (github.com/claudianeodev/ComfyUI-MOSS-Image-8GB), restart, and drag in the Fast workflow JSON.

How to save your own ComfyUI workflow as a CosyTemplate (and run it in Playground)

A common question is how to get a ComfyUI workflow into Playground. You save it as a CosyTemplate first. These are the steps we used to publish MOSS-Image 8GB:

Open your workflow in ComfyUI inside Promptus, and run it once to make sure it works.
Click PromptusAI in the ComfyUI toolbar, then Share. The CosyFlow Editor opens.
Give it a name and a description (under 255 characters). Avoid a colon in the name, because Promptus puts your username before the first colon.
Under Field Configuration, choose what people see. Each setting has a slider from 1 (Hidden) to 5 (Always). The main prompt is required and always shown.
Under Model Configuration, add a download link and folder for each model, so Promptus can fetch them for other people.
Under Example Media, choose a picture from your collection to show on the card.
Leave Visibility on Private to keep it to yourself, or switch it to Public to share it.
Switch on Distributed so other people get Install Locally for it. (Install Locally appears for templates that are both Distributed and Trusted.)
Click Save as New Cosyflow.
Open CosyTemplates and choose My CosyTemplates. New templates may not be marked trusted yet, so clear the Trusted filter under Properties if you don’t see yours.
Every template has a Share this template link on its page. Copy it to send to other people.

PromptusAI share dialog in ComfyUI inside Promptus — Step 2: PromptusAI › Share

Name and description in the CosyFlow Editor — Step 3: name and description

Model Configuration in the CosyFlow Editor — Step 5: Model Configuration

Visibility set to Public in the CosyFlow Editor — Step 7: Visibility set to Public

CosyTemplate page with the Share this template link — Step 11: the Share this template link

For workflows that use standard ComfyUI nodes, the template can then run in Playground. MOSS-Image 8GB uses its own node, so for now run it with Load in ComfyUI.

Tips for getting text right

Put the exact words in “quotation marks” in your prompt.
Keep the text short and big. Printed letters work better than cursive.
Draft at 512×512, then make your favourite at 768×768 (about 2.25× the time).
On an 8 GB computer, close other apps. Running out of memory is what slows it down most.

Chalkboard sign that says PROMPTUS, made with MOSS-Image on 8 GB — Chalkboard: PROMPTUS

Gig poster that says JUNO LIVE, FRIDAY 8PM, made with MOSS-Image — Poster: JUNO LIVE, FRIDAY 8PM

Honey jar with a label that says WILD HONEY, made with MOSS-Image — Label: WILD HONEY

Neon sign that says OPEN in English and Chinese, made with MOSS-Image — Neon: OPEN in English and Chinese

Measured results

Timed on a MacBook with an Apple A18 Pro and 8 GB of memory, using the prompt A chalkboard sign outside a cozy cafe that says “PROMPTUS”…

Model	Size	Steps	Time	Text correct?
Full MOSS-Image SFT (Q4_0)	768×768	20	30 min 27 s for the whole first run, including about 20 min of model downloads	Yes
Full MOSS-Image SFT (Q4_0)	512×512	16	10 min 44 s (text encoder 57 s, 16 steps at about 34 s each, decode 25 s)	Yes
4-step MOSS-Image (Q4_0)	512×512	4	2 min 16 s (text encoder 42 s, 4 steps in 68 s, decode 24 s)	Yes

Frequently Asked Questions

Can MOSS-Image run on 8GB?

Yes. With a 4-bit image model, a 3-bit text encoder and stable-diffusion.cpp, it runs on an 8 GB MacBook. In Promptus, use the MOSS-Image 8GB CosyTemplate.

How long does MOSS-Image take on an 8GB Mac?

On an Apple A18 Pro MacBook with 8 GB, a 512×512 picture took 2 minutes 16 seconds with the fast 4-step model, and 10 minutes 44 seconds with the full model. The first run also downloads the files: the 3.3 GB 4-step model took about 10 minutes on our connection.

Does it work on Windows with an 8GB graphics card?

The node downloads the official stable-diffusion.cpp build for Windows: CUDA for NVIDIA cards, Vulkan for others. We haven’t tested it on Windows yet.

How do I save a ComfyUI workflow to Playground in Promptus?

Save it as a CosyTemplate first. In ComfyUI inside Promptus, click PromptusAI › Share, fill in the CosyFlow Editor, and click Save as New Cosyflow. Then find it in CosyTemplates › My CosyTemplates.

Why does MOSS-Image spell better than other models?

It was trained for long, dense text in English and Chinese, and it reads prompts up to 4,096 tokens long. Putting the words in quotation marks helps.

Is MOSS-Image free for commercial use?

The weights are released under Apache-2.0, which allows commercial use. Check the licence for your own case.

How big is the download?

About 7.9 GB of model files plus a roughly 30 MB engine. Both download the first time you press Run.

Do I need to compile anything or use Terminal?

No. Install the CosyTemplate from CosyTemplates with Install Locally. The node downloads its engine and model files by itself the first time you press Run.

What’s the difference between the 4-step and the full model?

The 4-step model is the official SFT model with a 4-step speed-up merged in (Apache-2.0). It needs 4 steps at CFG 1 instead of 16–25 steps at CFG 3, so it’s much faster. Printed words came out right in our tests; cursive is less reliable.

Credits

MOSS-Image by OpenMOSS (Apache-2.0)
4-step merge by mischiefaudio, from akhaliq’s 4-step LoRA (Apache-2.0)
stable-diffusion.cpp by leejet (MIT)
Qwen3.5-9B GGUF by unsloth (Apache-2.0)
FLUX.2 VAE by Black Forest Labs (Apache-2.0)
MOSS-Image 8GB node and workflow (MIT), built and tested in Promptus: code on GitHub

Sources

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

An AI-generated adventurer character with a sword walking in Unity