Apple SiliconAvailable
M1 – M4 · Metal
Veizik for Apple Silicon →Solutions
What you can build on Veizik, honestly.
Use-cases drawn from what the engine does today — server, on-device, edge, multimodal and control. Each is marked Verified, Supported or On the roadmap, from the ledgers, never from a promise.
Server inference
Run your own checkpoints on your NVIDIA GPUs.
One native engine on NVIDIA CUDA, Ampere through Blackwell. The checkpoint you downloaded is the one it runs — no Python environment on the box, no converted model copy, no framework stack to patch.
Verified
No Python on the server
The engine is a compiled binary. It reads the published weights and executes them natively — nothing to pip-install and nothing to keep in sync.
Runs where the data is
Generation happens on your own hardware. Prompts, weights and outputs stay on the machine; only account, activation and downloads reach the network.
Measured, not projected
Every performance figure on the site is taken on named hardware with a released binary, bound to that build’s checksum.
Veizik for NVIDIA → NVIDIA results → Download →
Drawn natively by Veizik on an NVIDIA RTX 3090 — FLUX.1-schnell output, not a benchmark.
On-device inference
A model on the machine in front of you.
Language models run on Apple Silicon today, M1 through M4, as one Metal binary. On Android edge SoCs the engine is measured on the board and deployment is still pending — shown as roadmap, not as shipped.
Verified
Mobile deviceDeployment pending
On-device Android · measured on the board
Edge & mobile hardware →Edge AI
Inference on the board, not in a datacentre.
The engine is pure-native and dependency-free, so it runs on CPU edge boards — Raspberry Pi 5 and Zero 2 W — with no accelerator required. Language, on-device vision, image and video run on the board.
Supported
Runs on the CPUArm NEON native execution — no GPU, no NPU dependency to ship around.
Offline by constructionAfter activation a device runs with no connection; restarting ends the session lease.
Sealed, device-bound packagesEach build is a single sealed package bound to the device it was activated on.
Veizik on the edge → SoC hardware → Edge models →
Shown as what the engine draws on the board — output, not a benchmark. Each was rendered by Veizik natively on a Raspberry Pi, with no accelerator.
Multimodal
One runtime for language, image and video.
The same engine runs large and small language models, text-to-image and image-to-image, and text-to-video and image-to-video. One install, one command surface, one set of licence rules — across every class it supports.
Verified
Robotics & control
On-device control policies, weights included.
Reinforcement-learning control policies run as native on-device engines. These are the four policies that ship with their weights and are verified on hardware; Jetson control policies are on the roadmap.
Verified
Ships with the weights
The control policies are distributed with their weights (Stable-Baselines3 RL Zoo, MIT) — install and run, no separate download.
Deterministic execution
A control step is reproducible on the shipped build, the same discipline every published figure is held to.
Edge-GPU preview next
Jetson AGX Orin is a render-proven preview for generative models; Jetson control policies are on the roadmap and marked as such.
Enterprise deployment
Device-bound, local by design.
Activation binds a licence to each machine, the engine ships sealed and hardened, and generation stays on your hardware. Fleet-scale controls are in active development and are shown as roadmap, not as shipped.
Supported
Activated per deviceApprove a machine once in the browser; every published model then runs on it, offline, with no per-model licence.
A protected engineThe compiler and kernels ship sealed and hardened, so a deployment does not expose them.
AI transparency, built inEvery image or video build marks its output as AI-generated — in the file and in the pixels — with no switch to turn it off (EU AI Act Art. 50; Korea AI Framework Act Art. 31).
Fleet controls — on the roadmapMulti-GPU execution, a central licence server and an audit log are in development. We will say they ship when they do.
Security & licensing → Licensing → Talk to us about deployment →
Get started
Pick a platform and run a model.
Everything above runs from the same install. Start with your hardware, or read how the engine works.