QuadBays

AI on your own computer.
Set up in one step.

QuadBays sets up llama.cpp for your graphics card, downloads a model that fits, and opens a chat window. Your chats and documents stay on your machine. Windows, Mac and Linux.

Download QuadBays Buy Pro – $19

Servers

This computer (Vulkan)
Llama-3.2-3B-Instruct-Q4_K_M.gguf
Running
This computer (CUDA)
Llama-3.2-3B-Instruct-Q4_K_M.gguf
Stopped
You: What can you do on this computer?
Llama 3.2 3B: I run on your own graphics card, so nothing you type leaves

Your computer, your models

How QuadBays works

It sets up the engine

QuadBays looks at your graphics card and gets the matching llama.cpp from the llama.cpp project, checked against its published checksum. Vulkan runs on almost any card; NVIDIA cards also get CUDA beside it.

VulkanCUDAROCmMetal

It finds models that fit

Suggested models are listed by size against your card's memory. Search Hugging Face from inside the app. Downloads are checked, resume where they stopped, and a model you move is found again.

It chats, and remembers

Chats are saved on your computer and listed by date. The speed of every answer, and your card's temperature and power, are shown as it writes.

Pictures and videoPro

Make pictures and short videos from a description, on your own card, with stable-diffusion.cpp. Or insert a picture and ask a model that sees about it.

Up to four baysPro

Chats side by side, each on its own model or card, so you can compare answers. Build chats write files; Web chats can search the internet when you allow it.

Your documentsPro

Point Context at a folder of PDFs and the model reads them whole, then tells you which document to check. It reads; it never decides for you.

From install to first answer

  1. Install QuadBays. A setup program on Windows, a .dmg on a Mac, a package on Linux.
  2. Let it set up this computer. It names your graphics card and says what it will download, and how big, before it downloads anything.
  3. It downloads a small model and asks it a question. You finish setup with a server that has answered, not a list of settings.
  4. Chat. Add bigger models when you want them: each suggestion says whether it fits your card.

Version 1.0

Download

Windows

Windows 11, 64-bit. Tested with NVIDIA cards, from a GTX 1070 to an RTX 5080.

The engines need Microsoft's Visual C++ Redistributable. The installer offers it; installing it needs an administrator. On a work or school computer, ask your IT department.

Available at release

Mac with Apple silicon

M1 or newer. Runs on Metal. Two steps: open the .dmg, then run its installer.

Available at release

Intel Mac

Late 2015 or newer. Free edition only.

Available at release

Ubuntu & Debian

Ubuntu, Kubuntu, Debian, Linux Mint, Pop!_OS. A .deb package.

Available at release

Fedora & openSUSE

Fedora, Nobara, openSUSE Tumbleweed. An .rpm package. Aurora and Bazzite have their own one-click installer.

Available at release

Arch & CachyOS

Arch, CachyOS, Omarchy. Arch's own package, installed with one line or a double-click.

Available at release

Each download comes with Install and Troubleshooting.txt. Updating later: put the new download in your Downloads folder and QuadBays offers to install it.

One price, once

Free and Pro

Free

$0
  • Sets up llama.cpp for your card
  • Servers: start, stop and switch models
  • Models: suggestions, search and downloads
  • Chats, saved on your computer
  • The searchable Guide

The way to find out whether QuadBays runs well on your computer.

Pro

$19 once
  • Everything in Free
  • Pictures and video
  • Up to four chats side by side; Build and Web chats
  • Context: chat with folders of documents
  • The system prompt and generation settings
  • Cloud Providers and the Test Bench
  • External graphics cards (eGPU)
Buy Pro

No subscription and no account. The licence never expires and isn't tied to one computer. It arrives as a file by email. 14-day refund.

Common questions

Does anything leave my computer?▾
Your chats, models and documents stay on your computer. QuadBays goes online only when you ask it to: to get llama.cpp or stable-diffusion.cpp, to search or download models, to search the web in a Web chat, or to talk to a cloud provider you add. There are no accounts, no analytics and no tracking.
Which graphics cards work?▾
Vulkan runs on almost any card from AMD, NVIDIA or Intel, so every setup starts there. NVIDIA cards also get CUDA, and recent AMD cards on Linux can add ROCm, as a second server. Macs with Apple silicon use Metal. More graphics memory means bigger models.
Windows says VCRUNTIME140.dll was not found.▾
Microsoft's Visual C++ Redistributable is missing. Run the QuadBays installer again and answer Yes when it offers it, or install it from Microsoft (aka.ms/vs/17/release/vc_redist.x64.exe). It needs an administrator.
I'm on a work or school computer.▾
QuadBays installs for you alone, without an administrator. On Windows, its engines also need the Visual C++ Redistributable, which does need one. If you can't install it because of administrator restrictions, contact your IT department.
Do I need to know llama.cpp?▾
No. QuadBays fills in the settings for you. If you do know llama.cpp, every server setting is there to change, under its usual flag name.
Is llama.cpp included?▾
No. QuadBays is a separate app. Setup downloads llama.cpp from the llama.cpp project, or you can point QuadBays at a build of your own. Updating llama.cpp, or going back to an older build, is your choice.
How does the Pro licence work?▾
Buy it once and you get a licence file, QuadBays.licence, by email. Put it in your Downloads folder and QuadBays offers to add it (or use Help → Add licence). It has no expiry date, needs no account or internet check, and isn't locked to one computer.

Help

The Guide is built into QuadBays, with a search box. For anything else, write to [email protected] – a note and screenshots help most.

No accounts, no analytics, no tracking. QuadBays runs on your computer and keeps what you make there. It goes online only for the things you ask it to do.