Open the console / API guide / Data handling
Dedicated-service data handling
Implementation reviewed October 18, 2026. This page describes dedicated endpoints, their management API and managed inference routing. It is an operational notice, not new service terms or a request for training permission.
The separate Models API data policy and Models API terms, at inferoptimal.com/models/privacy and inferoptimal.com/models/terms, cover /inference/v1, including its disclosed 30-day content retention and training/evaluation permission. Those permissions are not extended to dedicated traffic by this page.
Where inference goes
Managed inference passes through InferOptimal’s Cloudflare gateway to the selected backend. Requests can include conversation text, images, tool definitions and tool results. Other dedicated endpoints can route directly to their serving infrastructure. Your deployment’s connection and preparation details identify its API and startup mode.
GLM immediate startup sends requests to Z.ai’s API while private GPUs load, then uses the private replica at the same URL and key. Private-only startup waits for the private replica instead. A private GPU does not mean the management service or gateway has no access to traffic; see the API guide for the mode you select.
What InferOptimal retains
The managed proxy’s analytics store numerical usage and outcomes, not prompts, completions, tool arguments or screenshot URLs. Operational diagnostics, deployment records and customer-selected runtimes are separate from those analytics.
| Data | Contents and retention |
|---|---|
| Managed request analytics | Backend, request outcome, token counts when reported, latency and retry counts. Up to 24 hourly bins; lifetime aggregate totals remain with the deployment record. The 24-hour chart window is not a deletion deadline for all metadata. |
| Engine and hardware metrics | Numerical counters, utilization and token-length distributions. Rolling metrics expire after 7 days without writes, machine windows after 1 day, history after 60 days and private numerical archives after 90 days. |
| Registered-runtime traffic shapes | Media dimensions, byte counts, durations, batch sizes and timing, not media or request text. A one-hour observation window expires after 7 days without updates. |
| Startup and failure diagnostics | A machine/configuration snapshot of up to 8 KiB stays with the deployment. Up to three engine-log excerpts of 32 KiB each expire after 90 days. Known request-log lines and secrets are filtered; arbitrary engine exception text is not guaranteed to be content-free. |
| Account, deployment and billing records | Account identity, configuration, lifecycle, usage summaries, credit holds and charges remain after Stop. These records have no single automatic deletion deadline in the current implementation. They are not covered by the short metrics windows above. |
Credentials and checkpoint files
Private Hugging Face read tokens and private-registry credentials supplied for deployment are kept in restricted control-plane storage with a 30-day expiry. Saved Lambda, Modal and GCP connection credentials also have a 30-day storage expiry. Stop does not immediately erase these saved credentials; disconnect or revoke access when it is no longer needed.
Platform API keys are stored as hashes with the expiry selected when created. Issued inference keys have a ten-minute retry-retrieval window; separate backend credentials can remain available to the service for routing and authorized restarts. Expiry or revocation of access is different from deletion of the associated account record.
The private-checkpoint staging path keeps the original HF token out of the native model process and uses bounded download access. Standalone AWS deployments use a separate credential path. Checkpoint files live on provider machines or storage; stopping inference is not a promise that every cached file or provider backup is immediately erased.
Z.ai startup processing
Z.ai’s published API Data Processing Addendum, section 4, states that API content is processed in real time without being saved on its servers, while other customer data can be retained for service and legal purposes. Its API terms state that API content is not used to develop or improve its services without explicit agreement; they also describe service, legal, policy-enforcement and abuse-prevention uses.
These are the provider’s published statements, checked October 18, 2026. They are not an independent audit or an InferOptimal guarantee of zero retention, a processing location, or account-specific contractual settings. Contact us before relying on a particular provider agreement or location requirement.
Other processors and your runtime
Cloudflare operates the gateway and control-plane stores; the selected compute provider hosts the serving machine. Hugging Face processes authenticated checkpoint downloads. Supabase or the connected account service handles sign-in, Stripe handles payments, and Resend sends lifecycle emails. Which services receive inference content depends on the route above; a billing or email integration does not by itself receive your conversation.
Customer containers, engine logging settings, uploaded files and persistent volumes can store additional data. This notice does not describe every custom runtime’s storage or a cloud provider’s backups. Numerical traffic summaries are not a stored replay of your conversations.
Account agreements and contact
Account sign-in, an individually agreed service contract and the separately accepted Models API policy are distinct. This page changes none of those agreements and does not add consent through sign-in. For the agreement applicable to your account, provider requirements, data access or deletion requests, contact team@velobyte.ai. Stop an endpoint to stop new inference; account-record deletion and provider-storage cleanup are separate requests.