The engine, the compiler, and where it runs.
One native runtime and one compiler across every model, architecture and backend. Here is how a run reaches your hardware, what each backend is, and how it installs — with links into the models you can run and the devices they run on.
Your checkpoint in. Native execution out.
The engine reads the published checkpoint as-is. The compiler lowers it to native code, and the run executes on the backend’s kernels on your own hardware — no Python, no framework, no numeric library, and nothing uploaded.
The checkpoint you already have is read as published; the compiler lowers it to native code and the engine runs it through the backend’s kernels on your hardware. Nothing is uploaded.
One engine, every backend.
Each execution backend changes only the kernels and the device. NVIDIA and Apple Silicon ship today; the edge CPU backend ships on single-board computers.
NVIDIA
AvailableCUDA execution on NVIDIA GPUs, on Linux. Language, image and video models run from the checkpoint you already have — no CUDA toolkit and no Python at inference time.
- Backend
- CUDA
- Platform
- Linux on x86-64
Apple Silicon
AvailableMetal execution on Apple Silicon Macs. Language models run natively on the GPU from your own checkpoint, with no Python environment and no converted model copy.
- Backend
- Metal
- Platform
- macOS on Apple Silicon (arm64)
Edge · CPU
AvailableThe same engine on a single-board computer’s CPU — a Raspberry Pi 5 and a 512 MB Zero 2 W — with no desktop GPU. A native, compile-free kernel path on the Arm cores does the work a GPU would.
- Backend
- Arm CPU · NEON
- Platform
- Linux on Arm — Raspberry Pi 5 (Cortex-A76) and Raspberry Pi Zero 2 W (Cortex-A53).
One engine, every backend — drawn natively by Veizik on an NVIDIA RTX 3090, a Jetson AGX Orin, a Raspberry Pi 5 and a Pi Zero 2 W. Output, not a benchmark. Powered by Stability AI.
Sealed, device-bound packages.
Each model ships as its own sealed, model-specific package: a single engine binary with no Python and no framework. A package is activated per device in the browser, and runs offline afterwards.
Device-bound, local by design.
Activation binds a licence to the machine, the first run leases it into memory, and generation happens on your hardware — prompts, weights and outputs never leave it. The compiler and kernels ship sealed.