webAI runs as a desktop app (via Tauri) and in the browser. The desktop app unlocks native features like MLX inference on Apple Silicon, local network discovery, and SQLite storage. The browser version works with WebGPU-capable browsers.
The desktop app requires macOS Tahoe (26) or later. The browser version has its own, less restrictive requirements — see Browser requirements below. | Platform | Support | Notes |
|---|
| macOS (Apple Silicon, Tahoe or later) | Full | Desktop app with MLX and llama.cpp inference backends |
| Windows | Coming soon | Desktop app in development |
| Web browser | Full | WebGPU inference only; some native features unavailable |
Browser requirements
WebGPU is required for in-browser AI inference. Supported browsers:
| Browser | Minimum version |
|---|
| Chrome / Chromium | 113+ |
| Microsoft Edge | 113+ |
| Safari | 17.4+ (macOS) |
| Firefox | Behind a feature flag (not recommended) |
If your browser does not support WebGPU, AI features will be unavailable. The rest of the app (spaces, chat, collaboration) still works.
Hardware
GPU
WebGPU inference requires a GPU with at least 1 GB of available video memory. The runtime requests high-performance GPU adapters and checks for:
- Minimum 8 storage buffers per shader stage
- Optional
shader-f16 (half-precision) for faster inference on supported GPUs
- Optional
subgroups for optimized GPU kernels
If optional features aren’t available, the runtime falls back to baseline kernels automatically.
Memory
Larger models require more system memory. General guidelines:
| Model tier | Parameters | Recommended system memory |
|---|
| Tier 1 | 2B | 4 GB |
| Tier 2 | 4B | 8 GB |
| Tier 3 | 12B | 16 GB |
| Tier 4 | 14B | 24 GB |
| Tier 5 | 32B | 64 GB |
| Tier 6 (MLX only) | 235B | 64+ GB Apple Silicon |
If model loading fails or is slow, try a smaller tier. The app will suggest an appropriate tier based on your device capabilities during onboarding.
Desktop vs. browser features
Some features are only available in the desktop app:
| Feature | Desktop | Browser |
|---|
| MLX inference (Apple Silicon) | Yes | No |
| llama.cpp inference | Yes | No |
| WebGPU inference | Yes | Yes |
| Local network discovery (mDNS) | Yes | No |
| SQLite storage | Yes | No (uses localStorage/IndexedDB) |
| Bluetooth device scanning | Yes | Partial (Web Bluetooth API) |
| File system access | Yes | Partial (File System Access API) |
| Spaces and collaboration | Yes | Yes |
| Voice chat | Yes | Yes |
| App uploads | Yes | Yes |
