NVIDIA CUDA
This lane carries the image and video work: FLUX.1, Stable Diffusion 3.5, LTX-Video, Wan and CogVideoX, on our own kernels, running here every day. It is not published yet, and this page is careful about the difference.
Running internally is not shipping
The CUDA engine is the same idea as the Metal one: kernels emitted by our own compiler, no framework beneath them, driven by the same command surface. Image and video families run on it here — FLUX.1, Stable Diffusion 3.5, LTX-Video, Wan and CogVideoX.
What is missing is not capability. It is the packaging work that turns a tree of build outputs into an artifact whose measured bytes and shipped bytes are provably the same file — which is exactly the property the Metal package had to earn before it was allowed out.
Until a CUDA package clears that, this site shows the lane as in preparation and offers no download.
The gate a lane must pass before it is published:
- The measured binaries and the shipped binaries are the same bytes.
- Every file in the archive is covered by a checksum list that verifies on a clean machine.
- Component compatibility is enforced before anything runs, not assumed.
- The catalogue's declared limits match what the runtime actually reports.
- Install and build paths contain no interpreter.
What the first CUDA package is expected to contain
Expected, not promised — the order is set by which parts clear the gate first, and the gate does not move for a date.
| Area | Target | State |
|---|---|---|
| Architectures | Ampere, Ada Lovelace, Hopper, Blackwell | In preparation |
| Image | FLUX.1, Stable Diffusion 3.5 | Runs internally |
| Video | LTX-Video, Wan 2.1 / 2.2, CogVideoX | Runs internally |
| Language | The families already open on Metal | Queued |
| Command surface | Identical verbs to the Metal lane | Built |
Engine entries on this lane
Loading…