Shipped — CPU and optional NPU.
Edge · CPU.
The 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.
The Edge · CPU backend.
What it is, in one view.
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.
This backend runs through the same engine and compiler as every other — only the kernels and the device change. The platform overview has the shared path.
Checkpoint in, native execution out.
The engine runs from a native Arm binary built by the same compiler; there is no Python and no numeric library on the board. It does a GPU’s job on four CPU cores through sub-byte precision chosen in-engine and NEON kernels.
What it does, and what it needs.
| Platform | Linux on Arm — Raspberry Pi 5 (Cortex-A76) and Raspberry Pi Zero 2 W (Cortex-A53). |
|---|---|
| Backend | CPU with Arm NEON; an optional Hailo-8 NPU on the Pi 5. |
| Model classes | Language and on-device vision run on the board today; image and video are render-proven and in preview. The models catalogue lists each one. |
| Features | Command-line generation; power and thermal governed for a render that finishes under load; see it running on a Pi CPU. |
| Install conditions | An arm64 Linux on the board and ordinary host tools; no GPU, no CUDA, no Python. Activate each device once in your browser. Model weights are not included — you supply the public checkpoint. |
What ships, and what does not yet.
Each tier carries the state it is actually in — shipped, a render-proven preview, or measured on-device with deployment pending.
An edge GPU (CUDA). Render-proven preview, not yet a download.
Confirmed on-device and measured on the phones; deployment is pending.