These terms govern the inference API at https://api.zephyr662.com.
They are written for wholesale and B2B customers — aggregators, platforms, and businesses that
buy inference capacity and resell or embed it. They are deliberately short: everything below is
a term we can actually honor.
These Terms are a contract between Zephyr 662 LLC (“we”, “us”) and the business entity that accesses the API (“Customer”, “you”). You accept them by issuing a request with an API key we have provisioned. The person accepting them must be authorized to bind the Customer.
The API is offered for business use only. It is not offered to consumers and is not intended for personal, household, or individual end-user account use.
If we and you have signed a separate written agreement covering the API, that agreement controls wherever it conflicts with these Terms.
We operate an OpenAI-compatible inference API. Concretely:
POST /v1/chat/completions endpoint accepting the OpenAI chat-completions
request shape, in streaming (server-sent events) and non-streaming modes, returning token
usage in both;GET /v1/models endpoint publishing the models we currently serve, with
per-token pricing in USD, context length, quantization, capacity, and datacenter location;The model catalog changes. We add, remove, re-quantize, and re-price models, and we may change
the hardware and inference engine behind any model. Model outputs are non-deterministic and will
differ across engine, quantization, and configuration changes even for identical inputs. We will
publish material catalog changes at /v1/models and, where a model is being retired,
aim to give advance notice — but no model is guaranteed to remain available.
Keys are credentials. Keep them secret, transmit them only over TLS, and rotate them if exposed. You are responsible for all usage under your keys, including usage by your own customers and end users, and for the fees that usage incurs. Tell us promptly at [email protected] if you believe a key has been compromised; we will revoke and reissue.
You may resell, resell-through, or embed the API in your own product and provide it to your own users. If you do, you must impose terms on those users at least as protective as Section 4 (Acceptable use), and you remain responsible to us for their conduct. We have no direct relationship with your users.
You may not use the API, or permit it to be used, to:
Because we retain no prompt or completion content (Section 7), we do not and cannot review your traffic after the fact to police this section. Enforcement is real-time and metadata-driven: rate limiting, key-scoped throttling, and suspension under Section 10. Enforcement may also follow a credible report, a lawful order, or your own disclosure. We rely on you to enforce this section against your users, since you can see traffic that we deliberately cannot.
Each key is subject to concurrency and rate limits, which we set and adjust based on available
capacity. Declared capacity in /v1/models reflects measured throughput at the time
of publication, not a reserved allocation.
When demand exceeds capacity, the API sheds load with an immediate HTTP 429
rather than queueing requests, so that a client can re-route without waiting. Backend degradation
is converted to a 429 at the gateway. This is a design choice about how we behave
under load; it is not a warranty that a particular class of response will never occur. Treat
429 as retryable with backoff, and do not retry 4xx responses other
than 408, 409, and 429.
The API is provided on an “as is” and “as available” basis. We make no commitment as to uptime, availability, latency, time-to-first-token, throughput, capacity, or continuity of any model, and we publish no service level agreement. We may perform maintenance, migrate hardware, or take capacity offline at any time. Nothing in our documentation, dashboards, marketing, or third-party measurements of our endpoint creates a service level commitment.
Do not deploy the API as a single point of failure in a system that requires guaranteed availability. Use a fallback provider.
Our data practices are set out in the Privacy Policy, which is
incorporated into these Terms. In summary: we never store your prompts or completions, we do not
train, fine-tune, or evaluate any model on them, and we retain per-request metadata — request
identifier (Inference-Id), API key label, model identifier, token counts, timing,
completion status, and computed cost — for billing, capacity management, and reliability, for up
to 12 months. The Privacy Policy is the controlling statement of our data practices; it governs
if this summary and it differ.
Because we retain no content, we cannot recover, replay, or produce your requests or responses — including for your own debugging, dispute, or compliance purposes, and including in response to a routine legal request. The Privacy Policy's legal-process section states the one forward-looking exception and the notice we would give. If you need a record of your traffic, log it on your side.
You are responsible for the content you send: for having the rights and any required consents or legal basis to send it, and for complying with the law that applies to it. Unless we have agreed otherwise in a signed writing, do not send data whose handling requires a specific contractual regime we have not signed up to (for example protected health information under HIPAA, or cardholder data under PCI DSS). A data processing addendum is available on request at [email protected].
As between you and us, you own your inputs, and we claim no ownership of the outputs generated for you. We grant you no rights in the underlying models beyond serving your requests.
The models we serve are third-party open-weight models distributed under their own licenses
and acceptable use policies. Those licenses run to you as the user of the outputs, and some of
them restrict fields of use, redistribution, or downstream training. Model identity, including
the Hugging Face model identifier and the quantization we serve, is published at
/v1/models so that you can determine which license applies. Complying with it is
your responsibility.
Outputs may be inaccurate, offensive, or unsuitable for your purpose, and may be similar or identical to outputs generated for other customers. Do not rely on outputs without review appropriate to the use.
Prices are per-token, in United States dollars, as published for each model in
/v1/models at the time a request is served. Separate input, output, and — where
applicable — cached-input and reasoning-token rates may apply. We may change prices on thirty
(30) days' notice by publication at /v1/models and to your billing contact; a price
change does not apply retroactively to usage already served.
Charges are computed from the token counts we report in the usage object of each
response and record against the request identifier returned in the Inference-Id
response header.
Unless a separate written agreement provides otherwise, we invoice monthly in arrears, in USD, for the preceding calendar month (UTC), itemized by model with prompt, completion, and total token counts and the corresponding amounts. Invoices are due within thirty (30) days of the invoice date. Billing contact: [email protected].
Dispute an invoice in good faith within thirty (30) days of the invoice date by notice to [email protected] identifying the disputed amounts; we will reconcile against our per-request ledger. Undisputed amounts remain due. Amounts not disputed within that window are final.
Overdue undisputed amounts may accrue interest at the lesser of 1.5% per month or the maximum permitted by law, and we may suspend service under Section 10 for non-payment. Fees are exclusive of taxes; you are responsible for all taxes on the transaction other than taxes on our net income. If you are required to withhold tax, you will gross up so that we receive the invoiced amount.
We may suspend or rate-limit your access, in whole or in part and with or without advance notice, if: you materially breach these Terms; we reasonably believe your traffic violates Section 4 or threatens the security, integrity, or capacity of the service or other customers; an invoice is overdue; or we are required to do so by law. Where practical we will notify you first and limit the suspension to what is necessary. Suspension does not relieve you of fees for usage already served.
Either party may terminate for convenience on thirty (30) days' written notice, and either party may terminate immediately for the other party's material breach that remains uncured ten (10) days after notice. We may terminate immediately if we cease offering the API.
On termination, keys are revoked and access ends. Fees for usage served before termination remain payable and are invoiced in the normal cycle. Sections 7 through 9 (as to accrued fees) and 11 through 14 survive.
To the maximum extent permitted by law, we disclaim all warranties, express, implied, and statutory, including merchantability, fitness for a particular purpose, title, non-infringement, and any warranty arising from course of dealing or usage of trade. We do not warrant that the API will be uninterrupted, timely, secure, or error-free, or that outputs will be accurate, complete, current, non-infringing, or fit for any purpose.
To the maximum extent permitted by law:
These limits do not apply to your payment obligations under Section 9, to your indemnity under Section 13, or to liability that cannot be limited by law.
You will defend, indemnify, and hold us harmless from any third-party claim, and any resulting loss, liability, damages, cost, and reasonable attorneys' fees, arising out of: your inputs; your use of outputs; your resale or provision of the API to your users and those users' conduct; your breach of Section 4 (Acceptable use) or Section 8 (model licenses); or your violation of law or of a third party's rights.
These Terms are governed by the laws of the Commonwealth of Puerto Rico, without regard to its conflict-of-laws rules. The parties submit to the exclusive jurisdiction of the courts located in San Juan, Puerto Rico for any dispute arising out of or relating to these Terms, and each waives any objection to venue there. The UN Convention on Contracts for the International Sale of Goods does not apply. Each party waives any right to a jury trial and to participate in a class or representative action.
We may update these Terms by posting a revised version at this URL with a new effective date. Material changes take effect thirty (30) days after posting; continued use after that is acceptance. If you do not accept a material change, stop using the API and terminate under Section 10.
To you: the billing or technical contact you have given us. To us, by email to [email protected] and, for any notice of breach, termination, indemnity claim, or legal process, also in writing to:
Zephyr 662 LLC
800 Ave Ponce De Leon Ste A, Unit 9121
San Juan, PR 00908
United States
Notices are effective on delivery. Service of legal process is governed by applicable law and may be made on our registered agent in Puerto Rico.
Neither party may assign these Terms without the other's consent, except to a successor in a merger or sale of substantially all assets.
Neither party is liable for failure to perform (other than payment) due to causes beyond its reasonable control, including power, network, datacenter, or upstream-supplier failure.
Neither party will use the other's name or marks publicly without prior written consent, except that a marketplace or aggregator may list us as a provider and we may state that we serve it.
If any provision is unenforceable, it is modified to the minimum extent necessary and the rest remains in effect. A failure to enforce is not a waiver.
These Terms and the Privacy Policy are the entire agreement between the parties on this subject and supersede prior discussions, except for a signed agreement as described in Section 1.
The parties are independent contractors. Nothing here creates a partnership, agency, joint venture, or employment relationship, and there are no third-party beneficiaries.