Platform Support
Linux is the first-class platform, with native macOS and Windows support.
Kaba client
Section titled “Kaba client”| Platform | Package |
|---|---|
| Linux | .deb, .rpm, AppImage |
| macOS | .dmg |
| Windows | Installer (.exe) |
[todo: confirm supported architectures (x64, arm64) and minimum OS versions for each client package.]
Linux notes
Section titled “Linux notes”- Both Wayland and X11 sessions are supported. Kaba detects the session and your GPUs and picks a rendering profile; if a profile fails to start with hardware compositing, the failure is recorded and the next profile is used.
- On machines with more than one GPU, the browser is pinned to the GPU driving the display, leaving the other free for model work.
- NVIDIA systems default to OpenGL rendering; set
KABA_NVIDIA_VULKAN=1to opt into Vulkan. - Hardware video decoding uses VA-API where available.
- The
Superkey is used for pane shortcuts. If your desktop environment claims those chords, rebind them under Settings → Keyboard.
Feature differences
Section titled “Feature differences”| Feature | Linux | macOS | Windows |
|---|---|---|---|
| Automatic updates | yes | [todo: confirm] | [todo: confirm] |
kabactl linked onto PATH | ~/.local/bin | printed instructions | user PATH entry |
| Passkeys with a hardware key that needs a PIN | limited | yes | yes, through the OS dialog |
| Media controls in the header | yes | [todo: confirm] | [todo: confirm] |
kabactl
Section titled “kabactl”Release binaries, from https://git.djcas9.com/kaba-labs/kabactl/releases, are named kabactl-<version>-<platform>.zip:
| Platform | Asset |
|---|---|
| Linux x86-64 | linux-x64 |
| Linux ARM64 | linux-arm64 |
| macOS Apple silicon | macos-arm64 |
| Windows x86-64 | windows-x64 |
Container images are published for linux/amd64 and linux/arm64.
kabactl self-update supports exactly these four platforms. Other targets can be built from source.
Service integration
Section titled “Service integration”| Linux | macOS | Windows | |
|---|---|---|---|
service start / stop (built-in daemon) | yes | yes | yes |
service install --systemd | yes | no | no |
service install --system | yes | [todo: confirm] | [todo: confirm] |
Tool sandbox
Section titled “Tool sandbox”| Platform | Runs commands with |
|---|---|
| Linux | An OCI runtime: gVisor runsc, youki, crun or runc. Images are pulled without Docker. |
| macOS, Windows | A Docker-compatible daemon. |
Inference and training back ends
Section titled “Inference and training back ends”| Back end | Selected with | Notes |
|---|---|---|
| wgpu (Vulkan, Metal, DirectX) | --device wgpu, or auto | In default builds. Used for training. |
| CPU | --device cpu | Always available as the fallback. |
| llama.cpp with Vulkan, Metal, CUDA or ROCm | build features | Chosen per release build. |
Local training needs a GPU. Inference works on CPU.
Hardware guidance
Section titled “Hardware guidance”| Use | Needs |
|---|---|
| Browser, terminal, files, mesh | Any machine that runs a modern browser comfortably. |
The small model (e2b) | About 2.8 GB on disk. |
The large model (e4b) | About 4.4 GB on disk. |
| Training adapters | A GPU. Otherwise train on a peer that has one. |
[todo: add minimum and recommended RAM and VRAM figures for each model, from measured numbers.]
The client sizes its own memory limits from the machine’s RAM, and limits the number of renderer processes on smaller machines.
Mobile
Section titled “Mobile”See mobile.