Skip to content
CozyGPU
Desktop app · Mac + Windows · v0.1.0 in progress

Open models.
No GPU required.

CozyGPU is a Mac and Windows app for open image and video models. It runs on your machine, inference in your own RunPod account, billed per second. The ease of a studio app, no node graph.

Download for MacComing soonDownload for WindowsComing soon
See how it works
01

Open models, current ones

Three open-weight models, chosen for output quality and a sane cost per run.

Each model runs the workflow its authors publish, on a GPU class that fits it, inside your RunPod account. No custom nodes to hunt down, no checkpoint folders to manage. New models arrive through a signed catalog without a reinstall.

Model details, settings and GPU classes
02

Three keys, five minutes

Paste a RunPod API key and an S3 key, optionally an xAI key, and generate.

There is nothing to install besides the app: no Python, no CUDA, no model downloads on your disk. CozyGPU creates one serverless endpoint per model in your RunPod account and scales it to zero when you stop.

Honest note: the first job on a new machine can take 10 to 30 minutes while RunPod pulls the model (about 10 for image models, up to 30 for video). After that, a warm worker answers in seconds.

Read the setup guide
  1. 1

    Install CozyGPU

    Download the Mac or Windows installer, open the app. Updates arrive on their own.

  2. 2

    Paste your keys

    A RunPod API key, a RunPod S3 API key, and, if you want prompt enhancement, an xAI key. They are stored encrypted on your machine.

  3. 3

    Pick a model and generate

    CozyGPU creates a serverless endpoint per model in your RunPod account, submits the job, and files the result in your library.

03

Your compute, your files

RunPod bills you per second of GPU time. Your library stays on your disk. Nothing goes to us.

You pay RunPod for GPU seconds and CozyGPU for the app. Workers idle out after ten seconds, so an afternoon of prompting costs cents, not a monthly GPU rental.

Images, videos, prompts, keys and settings are files and a database on your computer. Move them, back them up, delete them. We never see them.

What leaves your machine, exactly
Typical cost, your RunPod account
Z-Image Turbo image
about $0.003
Qwen Image 2.1 image
about $0.01
MiniMax H3, 5 s clip
$0.16 to $0.32
Idle
$0.00

Measured on the GPU classes the app selects; the library shows the real cost of each job.

Library6 of 1,284
  • A single wisp of white smoke curling upward on black
    Z-Image Turbo$0.003
  • Water droplets on black glass catching a warm light
    Z-Image Turbo$0.003
  • Three orange petals floating on still black water
    Qwen Image 2.1$0.01
  • Sheets of white paper curved into soft arcs on dark
    Z-Image Turbo$0.003
  • Molten glass glowing deep orange on dark
    Z-Image Turbo$0.003
  • Fine sand ripples lit by low warm side light
    Z-Image Turbo$0.003
04

Bring your own LoRAs

Drop a .safetensors file on the app. It shows up in the picker with a strength slider.

Name it, add trigger words, set a default strength. CozyGPU keeps the file on your machine, copies it to a small network volume in your RunPod account, and chains it into the workflow beside the built-in LoRAs.

Custom LoRAs work on all three models: Z-Image Turbo, Qwen Image 2.1 and MiniMax H3.

How LoRAs are stored and synced
A single wave of orange silk rising from black water under a soft moonseed 7777 · 9 steps · Z-Image Turbo
Pricing

One plan

CozyGPU license
$20per month

or $200 per year, two months free

  • All models, all modes, custom LoRAs
  • Mac (Apple silicon and Intel) and Windows installers
  • Auto-updates and new models through the catalog
  • One device per license
  • No trial. 14-day refund if it does not work for you.
Pricing details
Plus RunPod, billed to you by RunPod

GPU time is metered per second only while a job runs. You prepay credit on RunPod, set a spend limit if you like, and the app shows every job's measured cost.

Z-Image Turbo
about $0.003 per image
32 GB to 48 GB Ada or Ampere
Qwen Image 2.1
about $0.01 per image
48 GB Ada Pro (falls back to 32 GB Ada Pro)
MiniMax H3
$0.16 to $0.32 per 5 s clip
80 GB Ada Pro or 96 GB Blackwell

A month of 500 images and 20 short clips lands around $6 to $8 of RunPod credit. Heavy video use costs more; idle costs nothing.

FAQ

Questions

Something missing? opensourcegenai@gmail.com

Do I need a GPU?
No. CozyGPU runs on your machine, and inference runs in your own RunPod account. Any Mac from the last few years (Apple silicon or Intel) or a 64-bit Windows 10 or 11 PC is enough, including laptops with no discrete graphics at all.
What does RunPod cost?
RunPod bills your account per second of GPU time, only while a job runs. With the GPU classes CozyGPU picks, an image is about $0.003 on Z-Image Turbo and about $0.01 on Qwen Image 2.1; a five second MiniMax H3 clip is $0.16 to $0.32. You prepay credit on RunPod and can set a spend limit there. The app shows the measured cost of every job.
Is anything uploaded to you?
No. Prompts, input images and results move between your computer and your RunPod account. Prompt text goes to xAI only if you click Enhance and have added an xAI key. The only things that reach us are a license check and the update feed; the security notes list exactly what each one sends.
Which models are included?
Z-Image Turbo (text to image), Qwen Image 2.1 (text to image and multi-reference image edit) and MiniMax H3 (text to video and image to video, with audio). New models arrive through a signed catalog the app checks for, without a reinstall.
Does it run on Intel Macs?
Yes. There are separate Apple silicon and Intel builds, and because the heavy lifting happens on RunPod, an Intel Mac is not at a disadvantage.
Why does Windows show a SmartScreen warning?
Windows shows 'Windows protected your PC' for installers whose code-signing certificate has not yet built up a reputation with Microsoft, which is the case for a new app. The installer is signed; click 'More info' and then 'Run anyway', and compare the SHA-256 checksum on the download page if you want to be sure.
What is the refund policy?
There is no trial, so if CozyGPU does not work for you, email opensourcegenai@gmail.com within 14 days of your first payment and we refund the license. RunPod charges are billed by RunPod and are outside our control.
Can I use my own LoRAs?
Yes. Drop a .safetensors file on the app, pick the model it was trained for, give it a name and trigger words, and it appears in the LoRA picker with a strength slider. Custom LoRAs work on all three models: Z-Image Turbo, Qwen Image 2.1 and MiniMax H3.
Get CozyGPU

Open models on your Mac or PC, without a GPU.

Install, paste three keys, generate. $20 a month plus the RunPod seconds you actually use.

Download for MacComing soonDownload for WindowsComing soon