Where your data goes
Your prompts, input media and generated outputs run and stay on your hardware. This is a data-handling property: there is no upload path, so there is nothing to send. It is what the privacy policy describes.
VEIZIK for business
Understand what runs, connects and updates.
The runtime executes your models on your own hardware; your prompts, media and outputs stay on the device. The network is used for the account, device activation and downloads — and this page says exactly which is which. Every statement here links to the policy it comes from.
No certification mark or audit logo is shown here, because none is claimed. What is shown is the published policy text and the commands you can run yourself.
On your devicePrompts, input media, model weights and every generated output — they never leave the machine. There is no media-upload path in the runtime.
Over the networkAccount e-mail at sign-up, device activation and the licence lease, and downloads of the CLI and the sealed model packages. The control-plane is the
/api/* endpoints: sign-up, licensing, entitlement.Separate & optionalWebsite analytics (only with cookie consent) and disclosed crash / run diagnostics. The install-script request log records IP, user-agent and country — we disclose it rather than hide it.
Deployment surface
What connects, and how each part is verified.
Six areas of a deployment, each with the published policy or command behind it.
| Area | How it works — and where to read the policy |
|---|---|
| Dependency | One self-contained binary plus the GPU driver already on the machine. No Python, no deep-learning framework, no CUDA toolkit, no external numeric library. The published checkpoint is read in place, with no converted copy. Verify with ldd → |
| Licence binding | Compute is lease-gated. Activation binds the licence to a device — a readable device label, an opaque device id and when it was approved — and enforces the seat limit. Runs are local after activation. What is collected, and why → |
| Update & download verification | The CLI and each sealed model package are fetched over HTTPS and checked against a published SHA256SUMS before install; a checksum mismatch aborts the install. The client, engine distribution and the activation / update flow are in scope for our disclosure programme. Vulnerability disclosure policy → |
| Offline behaviour | Activation is a one-time browser approval and each model’s weights are pulled once, on demand. After that the run is local — an air-gapped or locked-down machine runs it the same as a laptop, with no per-run network call. Why it installs where Python will not → |
| Log retention | Enquiries are deleted once handled (kept up to three years only where Korean e-commerce law requires a complaint record). The install-script IP log is disclosed, and shortening what it stores is a tracked change. We do not store your IP against your account, keys, sessions or runtime telemetry. Retention & your rights → |
| Support & contact | Security reports go to security@veizik.com under the disclosure policy; privacy and data-subject requests go to support@veizik.com. The controller is LinkPick (brand VEIZIK), Republic of Korea. Make a data request → |
Two separate guarantees
Data handling and engine protection are not the same thing.
We keep them apart on purpose, so neither is read as the other.
How the engine is protected
The engine ships as a sealed package with anti-tamper integrity checks. That is a separate, software-protection property, and it is not a guarantee about your data. Our published threat model openly concedes that an attacker who already has root on the machine can reach the local runtime — a known, accepted limit, not a secret.
These are two different guarantees. “Your content never leaves the device” and “the engine resists tampering” are not the same security claim, and this page does not present them as one.
Have a deployment question?
If something here needs to fit a specific policy on your side, tell us the constraint and we will point you at the exact behaviour and its policy text. For how a workload is measured before deployment, see the evaluation process.
Discuss your deployment