This policy covers the inference API at https://api.zephyr662.com. It states
what our serving stack does — no more, no less. The claims below are enforced in code and
continuously tested.
We do not store your prompts or completions. Request and response content — messages, tool definitions, tool arguments, and generated text — exists only in process and GPU memory while your request is being served, and is released when the request completes.
Specifically, request and response content is not written to:
Transient in-memory caches used for performance — for example prefix and KV caches in accelerator memory — hold data only in volatile memory, are evicted in the normal course of serving, and are never written to persistent storage.
We do not train, fine-tune, evaluate, or improve any model on your prompts or completions, and we do not sell or share them — we could not, because we do not retain them.
For billing, capacity management, and reliability we record, per request: a request
identifier (Inference-Id), the API key label, the model identifier, token counts,
timing, completion status, and the computed cost. This metadata contains no message content.
Per-request records are retained for up to 12 months for invoicing and audit,
after which they are deleted or reduced to aggregate totals that identify no individual
request.
All traffic is encrypted in transit (TLS). Traffic between our gateway and GPU backends travels over an encrypted private tunnel. Access to production systems is key-based and limited to operations personnel.
Inference runs on rented GPU capacity located in Montreal, Canada, behind a gateway we operate on separate hosting. These providers execute inference in memory under the same zero-retention posture; none receive stored content, because none exists. A current list of infrastructure subprocessors is available on request at [email protected].
This policy describes what our software does. Two things sit outside it, and we would rather name them than imply otherwise. First, the physical machines running inference are operated by infrastructure providers whose personnel hold administrative access to their own hardware; we contract for capacity, and our stack writes no content to it, but we do not control their platforms. Second, anything your own client logs, caches, or forwards is under your control, not ours.
We do not retain content, so there is nothing to disclose in response to a routine legal request — the only records we can produce are the request metadata described above. If we were ever served with a valid legal order compelling preservation of future traffic, we would comply as required by law and — unless legally prohibited — notify the affected customer and update this policy's effective date before any change took effect.
This policy covers our systems only. Traffic reaching us has already passed through your own client and, if you reach us via a routing platform such as OpenRouter, through that platform — each with its own data policy. We cannot and do not make claims about those.
Material changes will be posted at this URL with an updated effective date. Questions or data requests: [email protected].
The entity responsible for the processing described here is:
Zephyr 662 LLC
800 Ave Ponce De Leon Ste A, Unit 9121
San Juan, PR 00908
United States
A per-surface breakdown of every path content could persist, and the control on each, is published at Zero-Retention Audit.