NVIDIA.
CUDA 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.
The NVIDIA 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 reads the published checkpoint, the compiler lowers it to native GPU code once per supported architecture, and the run executes those kernels directly on your card. There is no compile step at render time and no numeric library in the loop.
What it does, and what it needs.
| Platform | Linux on x86-64 |
|---|---|
| Backend | CUDA, through the driver only — the same build runs under the CUDA 12 and CUDA 13 driver series. |
| GPU architectures | Ampere, Ada Lovelace, Hopper, Blackwell — see the compute-capability table for each device. |
| Model classes | Language, image and video. The models catalogue lists each family offered on NVIDIA and its licence. |
| Features | Command-line generation, interactive chat and a local OpenAI-compatible HTTP server for text; frame-by-frame image and video rendering; a resident render server that serves request after request in one process; warm restore through the hypercache. |
| Install conditions | An NVIDIA GPU of compute capability 8.0 or newer, and an NVIDIA driver that provides libcuda.so.1. No CUDA toolkit and no Python — only the driver. The same build runs under both the CUDA 12 and CUDA 13 driver series; there is no separate CUDA-12 or CUDA-13 download. Linux x86-64, run natively (WSL is not supported). Activate each machine once in your browser (veizik activate); generation then runs offline on your GPU. Model weights are not included — you supply the public checkpoint. |