One native runtime across models, architectures and hardware — compiled, dependency-free, and run from the checkpoint you already have.
Six things the engine does that a framework stack does not.
No Python, no framework, no extra packages. One signed binary.
Kernels built for the device: CUDA on NVIDIA, Metal on Apple Silicon.
Runs the published weights. No converted model to depend on.
Language, image and video on one engine.
Per-architecture builds, from consumer GPUs to the data center and Apple Silicon.
Our compiler and kernels ship sealed — not as source.
The engine reads the published checkpoint and applies its own compilation and execution optimization as it loads.
The published safetensors, exactly as downloaded.
as publishedPrecision, compilation and scheduling chosen per model, at load time.
in-engineCode built for the GPU in front of it — CUDA or Metal.
per architectureGeneration runs locally. Prompts, weights and outputs stay put.
on-deviceEach platform names the devices its builds target. Availability follows each release, not the engine. The two platforms install differently.
CUDA · Linux on x86-64
| Device | Generation | Compute capability |
|---|---|---|
| NVIDIA GeForce RTX 3080 Ti | Ampere | 8.6 |
| GeForce RTX 3090 | Ampere | 8.6 |
| NVIDIA GeForce RTX 4090 | Ada Lovelace | 8.9 |
| NVIDIA GeForce RTX 5090 | Blackwell | 12.0 |
| NVIDIA RTX A5000 | Ampere | 8.6 |
| NVIDIA A40 | Ampere | 8.6 |
| NVIDIA A100 80GB PCIe | Ampere | 8.0 |
| NVIDIA A100-SXM4-80GB | Ampere | 8.0 |
| NVIDIA H100 80GB HBM3 | Hopper | 9.0 |
| NVIDIA RTX PRO 4500 Blackwell | Blackwell | 12.0 |
Metal · macOS on Apple Silicon (arm64)
| Device | Generation | — |
|---|---|---|
| Apple M4 Pro | M4 | — |
| Apple M1 Ultra | M1 | — |
| Apple M1 Max | M1 | — |
| Apple M2 Pro | M2 | — |
NVIDIA and Apple Silicon install differently.
One engine, two ways in. On NVIDIA, one veizik command installs each model engine by name. On Apple Silicon, nothing changes: the same signed disk image and the same commands.
Linux on x86-64 · one veizik command, model engines installed by name
One veizik command, model engines installed by name.
veizik command for Linux x86-64 once: curl -fsSL https://veizik.com/install.sh | sh.veizik install takes a short name or a model key and installs the package that carries it: qwen, qwen35, qwen36 or a Qwen model key such as qwen35_2b; wan, wan21, wan22 or a Wan model key such as wan21_t2v_1_3b (any case; qwen3.6 and wan2.2 work too). It downloads the package from its public GitHub release, refuses it unless its SHA-256 matches the published release, and unpacks it under ~/.local/state/veizik/engines/. veizik list shows what is available and installed; veizik uninstall (or veizik rm) removes a package.veizik activate, then veizik run qwen35_2b "…", veizik chat and veizik serve (each needs veizik activate once per device and the model's weights); veizik run qwen36_35b_a3b "…" --image FILE (or --video DIR) also reads an image or a short video.veizik activate once per device, then veizik run wan22_t2v_a14b "…" --out ./clip, or veizik serve wan22_t2v_a14b --port 8000 for a render API on 127.0.0.1 (POST /v1/videos/generations). veizik pull wan22 downloads the model's weights and prepares the video decoder automatically; then render.How the NVIDIA release installs
One veizik command; each model engine is installed by name.
# the veizik command for Linux x86-64 $ curl -fsSL https://veizik.com/install.sh | sh # if the installer asks you to put ~/.local/bin on your PATH $ export PATH="$HOME/.local/bin:$PATH" $ veizik list # Qwen (language): install by qwen, qwen35, qwen36 or a full model key (qwen35_0p8b, qwen35_2b, qwen35_4b, qwen35_9b, qwen36_35b_a3b) $ veizik install qwen36 # have the Qwen3.5-0.8B checkpoint on this machine first: activation ends with a short check generation that uses it $ export VZ_WEIGHTS=/data/models $ veizik activate $ veizik run qwen36_35b_a3b "The capital of France is" --tokens 32 $ veizik chat qwen35_2b $ veizik serve qwen35_2b --weights /data/models/Qwen3.5-2B --port 8000 # Wan (video): install by wan, wan21, wan22 or a full model key (wan21_t2v_1_3b, wan21_t2v_14b, wan22_t2v_a14b, wan22_i2v_a14b) $ veizik install wan22 $ veizik activate # once per device; for the Wan package too $ veizik pull wan22 # fetches the Wan weights and prepares the video decoder $ veizik run wan22_t2v_a14b "<your prompt>" --out ./clip $ veizik serve wan22_t2v_a14b --port 8000 # render API: POST /v1/videos/generations
macOS 13 or later on M1 or newer · signed disk image
Veizik for Apple Silicon, as it ships today
install.sh from the volume it opens.veizik pull; there are no engine packages to add.$ ./install.sh $ veizik activate $ veizik pull <model> $ veizik run <model> "…" $ veizik chat <model> $ veizik serve <model>
Sign in once and approve the machine in your browser. It receives a licence bound to that device, and every published model then runs on it — no separate per-model licence. Nothing else needs to be configured.
The first run of a session leases the licence key into memory. Runs that reuse it need no connection, and restarting the computer ends the window.
Generation happens on your hardware. Prompts, weights and outputs never leave the machine; only account, activation and downloads reach the network.
The compiler and kernels ship sealed and hardened, so a deployment does not expose them.
Veizik follows the transparency rules of the EU AI Act (Art. 50) and Korea’s AI Framework Act (Art. 31). It says up front that it is generative AI, and every image or video build marks its output as AI-generated — in the file and in the pixels — with no switch to turn that off. How each rule is met →
Kernels, precision and memory layout tuned per GPU architecture rather than one generic path.
The compiler and kernels ship sealed and hardened, so your deployment does not expose them.
Sealed, model-specific engine packages with device-bound licensing. What you qualified is exactly what runs.
Generation runs locally; activation is once per device and runs need no connection.
A single engine across language, image and video means one thing to qualify, update and support.
One activation covers every published model on the machine — free for individual and non-commercial use, with no separate per-model licence. Enterprise and commercial use needs a separate licence; talk to us.
A five-second 1280×720 video render survived two host crashes and three other stops, resuming each time from its last step — and a resumed render is byte-for-byte the render that never stopped.