Its own stack, top to bottom.
Most local-inference setups are a tower of other people’s software — a Python environment, a deep-learning framework, a numeric library, a toolkit — and any one of them can break the run. veizik is one binary that talks to the driver and nothing else. You can check that yourself on the file you download.
Weights in, native execution out.
What the engine sits on, and — just as important — what it doesn’t.
What is not in the stack
- No Python runtime. Nothing to install, pin or import.
- No deep-learning framework. No PyTorch, no diffusion or transformer runtime to install alongside it.
- No CUDA toolkit. The NVIDIA driver is all the GPU side needs.
- No external numeric library. No cuBLAS, no cuDNN.
- No model conversion. The published checkpoint is read in place — no second copy in a private format.
The published weights go in; native execution comes out. Between them is one binary and the driver that is already on the machine.
Check it on your own download.
Independence you can confirm in three commands, not a claim you have to believe.
You do not have to take this on trust. Three commands on the binary you downloaded show that it stands alone — one file, and nothing from a framework or a numeric library linked into it.
# it is one file, not a tree of packages $ ls -lh veizik $ file ./veizik # the engine reaches the GPU through the driver — not through a framework, # a numeric library, or a Python runtime $ ldd ./veizik | grep -Ei "python|torch|cublas|cudnn" || echo "none of those are linked"
Run them against the file from Releases. The commands are standard tools; what they print is a property of the binary, not of this page.
Fewer moving parts, fewer ways to fail.
- It installs where Python won’t. No environment to build, no versions to reconcile — the binary and the driver are the whole dependency list, so an air-gapped or locked-down machine runs it the same as a laptop.
- Nothing underneath can drift. There is no framework release to track and no package tree to break between one run and the next; the engine is the same bytes every time.
- One engine, three kinds of model. The same engine runs language, image and video — delivered as a sealed package per model — so you learn one command and one install, not a stack per task.
- It is small. The Apple Silicon build is about 6 MB; each NVIDIA model package is a few megabytes, with its exact size and checksum on the download page.