Download Qwirbel
The packages are ready in your account. Sign in once – you then stay signed in.
The same app on Windows, Linux and macOS – same features, same key.
The downloads live in your account
Sign in once with the email address from your purchase – you then stay signed in until you sign out yourself. Your license keys are in the account too.
What your hardware needs to do
Qwirbel itself is small. Space and power are needed by the models you run with it – so here are the real numbers instead of a rule of thumb.
| GPU memory | What that gets you | Tier |
|---|---|---|
| 6 GB | Small language models (4B class) and images at lower resolutions. 8B models only work with offloading. | Entry |
| 8 GB | 8B models such as qwen3:8b (5.2 GB) or llama3.1:8b (4.9 GB) fit entirely on the card. SDXL-class images. | Solid |
| 12 GB | gemma3:12b (8.1 GB) and qwen3:14b (9.0 GB) fit on the card together with context. Flux-class images. | Recommended |
| 16 GB | Additionally video workflows (WAN/LTX) and longer contexts without constant offloading. | For media |
| 24 GB+ | Large models (27B/30B class), video without pressure, and several agents at once on one card. | Everything |
Model sizes come from the model list inside the program. Actual GPU usage is slightly higher because the conversation context has to fit as well. If GPU memory runs short, Qwirbel offloads parts into system RAM – it keeps working, just slower.
RAM & swap
16 GB is the entry point, 32 GB is the comfortable size, 64 GB stays relaxed even with several models.
Why: while loading, a model briefly sits in memory twice, and image and video workflows keep several parts around at once. If space runs out, the run aborts. Swap makes nothing faster – it prevents the abort.
Disk space
The program itself is a download of roughly 25 MB. The space goes into models: 5–9 GB per language model, image and video models are larger.
Plan for 100 GB free to start, and considerably more if you want to keep several models side by side. An SSD is strongly recommended – on a spinning disk, loading alone takes noticeably longer.
Graphics card
NVIDIA cards are the most well-trodden path. AMD cards work too – for image models via Vulkan, see the note below. Qwirbel also runs without a graphics card: language models then use the processor, noticeably slower, but they do work.
If your card is small
Two honest options: pick a smaller model – or add an API key, then the provider does the work and your card stays free. Both are a switch inside the program, and “Fully local” turns providers back off at any time.
How that worksAlso here: image AI for AMD & Vulkan
freeKlecks is our own interface for image models – built like ComfyUI but on Vulkan, so AMD cards simply join in. Since v2.6 it also runs language models. Free, no account and no licence – unpack it and place it next to the Qwirbel folder.
About this page
Purchase & pricing
Download & account
Activation & devices
Features
Support & community
I’m not online right now. Write anyway – I’ll read it and get back to you.