Terms of Service

Version 1.0. First published 5 August 2026.

1. Who we are

The Service is operated by CYANKIWI LTD, a private limited company registered in England and Wales with company number 16974010, whose registered office is at 64 Nile Street, International House, London N1 7SR, United Kingdom. In these terms we are “cyankiwi”, “we” or “us”, and you are “you”.

Report misuse of the Service to abuse@cyan.kiwi. Contracts, subpoenas, formal legal notices and law enforcement process go to legal@cyan.kiwi. Data protection questions, rights requests and complaints go to privacy@cyan.kiwi. Help with your account and general enquiries go to support@cyan.kiwi, and security reports go to security@cyan.kiwi. Post goes to the registered office above.

2. Definitions

Account is the account you open to use the Service. Credits are the prepaid balance recorded against your Account, redeemable only against the Service. Customer Content is what you send us: requests, prompts, files, model weights and datasets. Output is what a model returns in response. Model catalog is the machine-readable list of the models we serve at api.cyan.kiwi/v1/models, presented readably on the model catalog page. Service is the API at api.cyan.kiwi, the dashboard and site at cyan.kiwi, the model catalog, and anything else we agree in writing to provide under these terms. A working day is a day other than a Saturday, Sunday or English bank holiday.

3. The documents that make up this agreement

These terms are the agreement between us. Two documents are incorporated into them:

• The Acceptable Use Policy always forms part of these terms. It binds you and everyone who uses the Service through you, and you do not accept it separately: agreeing to these terms is agreeing to it.
• The Data Processing Addendum forms part of these terms automatically whenever we process personal data on your behalf, that is, whenever Customer Content you send us contains personal data. You do not need to sign it or ask for it: if that processing happens, the Addendum already governs it. Its schedule of authorised sub-processors is the sub-processor register, which is part of the Addendum, not a separately accepted document.

Per-model terms are incorporated at the point of use: when you select or use a model, the upstream licence and any conditions identified for that model on the model catalog apply to your use of that model and its Output (clause 6).

Two documents sit deliberately outside the agreement:

• The Privacy Notice is a transparency notice describing how we process personal data as a controller. You acknowledge it; you do not agree to it. It is not a contract term, and nothing in it is, or asks for, consent to processing. Our data protection obligations to you live in these terms and in the Data Processing Addendum.
• The Trust Center describes our security posture and operational practices so you can evaluate us. It is informational, not contractual, and it cannot reduce the binding security measures in the Data Processing Addendum (Annex III there).

Order of precedence. If these documents conflict: the Data Processing Addendum governs on the processing of personal data; a model’s upstream licence governs on what you may do with that model and its Output; otherwise these terms govern, then the Acceptable Use Policy. Clause 23 sets out how these documents change and is the only change mechanism, except that sub-processor changes run on the notice and objection process in the Data Processing Addendum.

By opening an Account or using the Service, you agree to these terms and the documents incorporated into them.

4. Eligibility, Accounts and API keys

The Service is sold for business and professional use: putting model inference inside your own products, tools and workflows. By opening an Account you confirm that you are at least 18 and are acting in the course of a business, trade or profession, that your Account information is accurate, and that you can enter into these terms.

You must be 18 or over to use the Service. If you are under 18 you may not open an Account, use an API key, or use the Service through anyone else’s Account. We may close an Account we believe belongs to someone under 18.

You are responsible for your API keys and for everything done with them. Keep them secret. Rotate a key from the dashboard immediately if you think it has been exposed. We store keys only as hashes, so we cannot recover one for you.

You may close your Account at any time by emailing privacy@cyan.kiwi from the address it uses. The Privacy Notice describes the process, and the statutory deletion rights it describes apply regardless of these terms.

5. The Service

The Service is a developer API for large language model inference, compatible with the OpenAI API format, sold on prepaid Credits. Model availability, pricing and declared capabilities are published on the model catalog.

Responses returned by the API are generated by AI models, not by people.

We relay your requests and the responses substantially unchanged, subject to a small number of documented protocol behaviours described in the API documentation. Parameters a model does not support may be dropped rather than rejected. Inference runs on our own inference stack, on serverless GPU infrastructure from the providers identified on the sub-processor register.

We do not currently operate content classifiers or filters in the serving path, and Output does not currently carry watermarks or provenance metadata (clause 17 states where we stand on marking). That is a description of how the Service runs today, not a promise that it will always run that way: we may introduce targeted safeguards where the law requires them, and we will describe any such change here and in the Acceptable Use Policy before it takes effect.

6. Models: licences, retirement and preview

The model catalog is the canonical list. Which models we serve, who publishes them, what they cost and what licence governs your use of them are published on the model catalog. The catalog is generated from our deployment configuration, so the published list reflects what is deployed, and our deployment process requires a licence record for every model before it is served.

Upstream licences flow down to you. We serve open-weight models trained and published by others. For each model the catalog identifies the upstream repository, the licence, a link to the licence text, and any operative condition that falls on you rather than on us. You must comply with the upstream licence for any model you use, in your use of both the model and its Output; it is incorporated into these terms for that model, and clause 11 explains why it wins where it and our assignment disagree. We record each licence at intake by reading the upstream repository’s own licence file rather than trusting a hub tag, and we write an intake record for every model before it is served; the Trust Center describes that practice.

Generally available and preview. A model is generally available when it is listed on the model catalog and not marked preview. A model is preview when its catalog entry carries is_ready: false. Preview models carry no availability or retirement commitment and may change or be withdrawn at any time; do not build production dependencies on them.

Retirement. We may add, change or retire models at any time. Retirements of generally available models are announced on the model catalog; we aim to give reasonable notice, without committing to a fixed period. Discontinuing the whole Service is governed by clause 22.

7. Availability, support and rate limits

We do not currently offer a service level agreement.

What remains is the law’s own floor, which we do not exclude: where a statute implies into the agreement a term that the Service is performed with reasonable care and skill, that term applies, and your remedies for its breach are the ones the law gives you (clause 19). We monitor availability continuously for our own operations, but that monitoring is an internal practice, not a contractual commitment.

Support is best effort, by email to support@cyan.kiwi; we do not commit to response or resolution times. Security reports go to security@cyan.kiwi under the reporting rules in the Trust Center.

Rate limits. Each API key carries a requests-per-minute and a tokens-per-minute limit, returned as 429 when exceeded. Capacity is shared, so under load requests may also be shed with a 429 rather than queued.

8. Acceptable use

You must comply with the Acceptable Use Policy, and so must anyone using the Service through you. Its section 6 describes what we can and cannot detect; do not read the prohibitions as a promise that we are watching.

9. Giving others access

If you resell, aggregate or otherwise make the Service available to others, you must impose restrictions on them at least as protective as the Acceptable Use Policy and these terms, pass on the transparency obligations in clause 17, remain responsible for what they do, and not knowingly give access to anyone we could not lawfully serve ourselves under clause 16. If we reasonably need information about that use to investigate a report or meet a legal obligation, you will provide it within 10 working days, or sooner if an authority sets a shorter deadline.

10. Reporting misuse and legal notices

Report misuse to abuse@cyan.kiwi. Send contracts, subpoenas, formal legal notices and law enforcement process to legal@cyan.kiwi. We review misuse reports and act on what we find, act on notices from competent authorities within the time the law allows, and cooperate with lawful requests. Section 5 of the Acceptable Use Policy has the detail.

11. Your content, Output and intellectual property

You keep all rights in Customer Content. As between you and us, you own the Output generated for you, and we assign you any right, title and interest we may have in it.

Two honest limits on that. Output is probabilistic and may not be unique: similar or identical Output can be generated for other users, and we make no claim that Output is protectable or that you are the only person who will receive it. And our assignment can only pass on what we hold: we serve models published by others under their own licences, and the chain of rights runs from the model publisher, to us, to you, so your use of Output remains subject to that model’s upstream licence, which is identified on the model catalog (clause 6) and wins where the two disagree, because it is the one with the rights behind it.

We do not restrict you from using Output to develop, evaluate or train your own models. That position is subject to the upstream licence: if a publisher’s licence restricts it, that restriction reaches you through the catalog entry and we cannot waive it on the publisher’s behalf.

You grant us the right to process Customer Content only to provide the Service. Nothing more. We do not use it to train or improve any model, and there is no setting that changes that.

Do not send us regulated data we have not agreed to handle: no protected health information (we do not offer a business associate agreement), no payment card data, and nothing else whose processing needs a compliance regime we have not agreed to in writing.

12. Models we publish on Hugging Face

We publish quantized versions of open models under the cyankiwi organisation on Hugging Face. Those are free downloads, not part of the Service, and downloading one does not make you our customer.

Your licence comes from the original publisher, not from us. Quantizing a model does not create a new licence and we do not purport to grant one. Each published repository identifies the model it derives from, and the upstream licence governs your use, including any naming, attribution, notice or acceptable-use condition it carries. Some of the families we publish (the Llama releases in particular) impose specific naming and attribution requirements on anyone who redistributes them, and some carry bespoke licences that have to be read rather than assumed from a tag. We keep a register of what each published repository is and of the licence text and attribution its upstream requires. If you find a repository where that is wrong or missing, tell us at legal@cyan.kiwi and we will fix it.

What we warrant about a published model: very little, deliberately. These are free artifacts provided as is. We state the quantization method on each model card and we test what we publish, but quantization is lossy and we make no warranty of fitness, accuracy or equivalence to the original weights. Evaluate before you deploy.

Rightsholders and takedowns. If you hold rights in material you believe is in one of our published models or in our published calibration dataset, write to legal@cyan.kiwi. If you want personal data removed from either, write to privacy@cyan.kiwi. We will respond, and we will act where the claim is made out.

13. Confidentiality

Each of us may receive information from the other that is marked confidential or that a reasonable person would understand to be confidential. Each of us will use the other’s confidential information only for this agreement, protect it with at least reasonable care, and not disclose it except to people who need it and are under equivalent obligations. This does not apply to information that is public through no fault of the recipient, was already known to it, is independently developed, or must be disclosed by law. Where the law compels disclosure, the recipient will give notice first if it lawfully can. These obligations last three years after the agreement ends, and indefinitely for anything that is a trade secret.

14. Data protection

Two kinds of personal data move through the Service, and the documents split along that line.

Account Data (the account, billing, usage and log data we collect to run the Service) is data we control. The Privacy Notice describes what we hold, why, for how long, and your statutory rights over it. It is a transparency notice, not part of this agreement (clause 3), and your rights over that data come from data protection law, not from contract.

Customer Content may contain personal data other people entrusted to you. Where it does, you are the controller or a processor acting for another controller, and we are your processor or sub-processor on the terms of the Data Processing Addendum, which applies automatically (clause 3). We do not assume your counterparty is the controller, and the Addendum is written for the intermediary case as well as the direct one.

We do not store the content of your requests or the responses. The binding statement of that commitment, including its two narrow exceptions, is clause 3.2 of the Data Processing Addendum, and the Addendum controls on the handling of personal data in Customer Content.

15. Credits, fees and taxes

Usage is prepaid. Credits are purchased in advance and consumed by inference at the per-token prices published for each model. When your balance runs out, requests are rejected until you top up.

Prices are exclusive of tax. You are responsible for VAT, sales and similar taxes on your purchases, other than taxes on our income. We may change prices at any time by publishing the new prices on the model catalog; the new prices apply to requests served after publication. Credits are a prepaid balance in dollars, not a quantity of tokens: consumption is metered at the per-token prices in force when each request is served, so after an increase an existing balance buys fewer tokens than it did before. There is no price grandfathering.

Credits have no cash value beyond the return rights below, are not transferable between Accounts, and are not a deposit or e-money: they are a prepayment for our own services.

Refunds. Credits you have consumed are not refundable. Credits you have bought and not consumed are refundable on request: the billing page has a self-serve refund that returns your unconsumed purchased Credits to the original payment method or methods, most recent purchase first. That is the only destination we send money to. We refund the full unconsumed amount, and we absorb the payment processing fees rather than deducting them. Your bank controls how long the money takes to arrive. To limit abuse there is a small minimum, one self-serve refund in any seven-day period, and refunds pause while an auto-reload payment is still settling. Consumption between your request and its execution reduces the refund accordingly. Credits we granted rather than sold are not refundable in any case. If we terminate for a reason other than your breach, or discontinue the Service, we return your unconsumed purchased Credits without you needing to ask. And if a statute that cannot be excluded by agreement gives you a cancellation or refund right (clause 24), that statute applies as written.

Saved payment methods and auto-reload. When you buy Credits at checkout, the payment method you use is saved with our payment processor for future use; we store only a reference to it, never the card details, and the billing page shows which methods are saved. Saved methods serve two purposes. They are where refunds go. And they power auto-reload: enabling it authorises us to charge the payment method you most recently paid with, without further confirmation, for your chosen reload amount whenever your balance falls below your chosen threshold. You can change the amounts or turn auto-reload off on the billing page at any time, and you can remove a saved payment method on the payment processor’s own management page; turning auto-reload off stops future automatic charges but does not undo one already in progress, and removing the payment method auto-reload would charge turns auto-reload off. If an automatic charge fails repeatedly, or needs a confirmation from your bank that an automatic charge cannot provide, we stop trying until you complete a payment or re-save your settings. Credits bought automatically are ordinary purchased Credits, including for refunds.

Chargebacks and refunds we make reduce your balance accordingly. If that leaves you below what you have already consumed, the Service stops until you top up.

Sanctions. Where sanctions or export law requires it, we may freeze a balance, decline a purchase and withhold a refund or other payment, without liability. Prepaid Credits and access to compute are economic resources for sanctions purposes, so this is not a formality.

16. Export control and sanctions

You must not use the Service in breach of export control or sanctions law.

You represent, on a continuing basis and each time you use the Service, that neither you, nor anyone who owns 50% or more of you or otherwise controls you, nor anyone you are acting for, is: designated on the UK Consolidated List, the EU consolidated list, a United Nations list, or a list maintained by the US Office of Foreign Assets Control or Bureau of Industry and Security; or located in, ordinarily resident in, or organised under the laws of a jurisdiction subject to comprehensive sanctions. You will tell us promptly if that stops being true.

Jurisdictions we do not serve currently include Cuba, Iran, North Korea, Syria, and the Crimea, Donetsk, Luhansk, Zaporizhzhia and Kherson regions of Ukraine. UK measures also restrict providing IT and software services to persons connected with Russia and Belarus, and we do not serve them.

The United States reaches this Service. Our inference providers, RunPod, Inc. and Modal Labs, Inc., are United States companies, our platform agreement is with Google Cloud EMEA Limited, an Irish entity that engages its United States parent as a sub-processor, the technology we run is largely of US origin, and our infrastructure sits in the United States. US export control and sanctions law therefore applies alongside UK and EU law. The sub-processor register records who each provider is and where processing happens. You must not export, re-export or transfer the Service or anything derived from it in breach of those laws, and you must not use the Service in connection with a chemical, biological, nuclear or missile end use.

If you become aware that any use may be intended for a sanctioned destination or end use, tell us promptly. We may screen Accounts and payments against sanctions lists, ask you for information reasonably needed for sanctions or export compliance, and refuse, suspend or terminate service immediately and without notice where the law requires it or where we reasonably suspect a breach. Once a competent authority informs us in writing that a licence is required, we have no discretion to continue without one. We apply the same rule on our own knowledge: where we know, or have grounds to suspect, a controlled end use, we act as if we had been notified, whether or not we have been.

17. AI Act roles and transparency

Our position. We are a provider of AI systems and a downstream provider: we place an inference service on the market under our own name, built on models trained by others. We are not a provider of a general-purpose AI model. We also publish quantized versions of open models under our own name (clause 12), and that publication is why this conclusion needs reasons rather than assertion: quantization is an efficiency optimisation of an existing model. It involves no training and does not approach the modification threshold at which a downstream modifier becomes the provider of a general-purpose AI model. That is a per-model assessment made at intake, not a blanket claim, and we keep it under review; if it changed for any model, the model-level obligations would attach to us for that model and we would say so here.

Where we stand on marking. Article 50(2) of the EU AI Act requires providers of systems that generate synthetic content to mark outputs in a machine-readable format, and applies from 2 August 2026. That duty falls on us for the API, and we are stating plainly that Output does not currently carry any machine-readable provenance marking. The Service was placed on the market before 2 August 2026, so the marking duty benefits from the four-month transitional period (to 2 December 2026) in Article 111(4) of the AI Act, inserted by Regulation (EU) 2026/1744, and we intend to implement marking before it ends. That transition covers the Article 50(2) marking duty only: the disclosure duties in Articles 50(1) and 50(4) apply from 2 August 2026 with no relief. Clause 5 carries our side of that disclosure, and the obligations below are yours from that date. If you need provenance today you must add it yourself.

Your obligations. If your product lets people interact with model Output, you must ensure they are informed they are interacting with an AI system unless that is obvious from the context, and you must pass that requirement on to anyone you give access to. If you publish deep fakes or AI-generated text informing the public on a matter of public interest without human editorial responsibility, you must disclose it. If you build the model into a product of your own, you are likely the provider of that product for AI Act purposes, and the marking duty is yours as well as ours, and ours does not discharge yours. If markers are present in an Output, do not strip them.

18. Optimisation, training and evaluation services

Where we separately agree to quantize a model, train a draft or speculative decoding model, or run evaluations for you, this clause applies unless we agree otherwise in writing.

We screen this work before we accept it. Before we take on an engagement under this clause we assess whether the artifact we are asked to produce is designed or optimised for a purpose Tier 1 of the Acceptable Use Policy prohibits, and we record that assessment before work starts, the same way we record model intake. We will refuse, and will stop, any work we assess as being for such a purpose, and we may decline to explain the assessment in detail where doing so would help someone route around it.

You keep all rights in the model weights and datasets you give us, and you warrant that you have the rights to give them to us for that purpose. You own the resulting artifact (the quantized or trained weights and the evaluation results), and we assign you any rights we have in it.

We keep our own methods, tooling and recipes. Improvements of general application to those remain ours; refinements specific to your model or your data are yours and we will not reuse them. We will not use your weights or your datasets to improve our general pipeline, to build any model we publish, or for any purpose other than performing the work you asked for, unless you opt in separately and in writing.

When the work is finished we will delete your weights and datasets, or return them, as you choose, without undue delay. Quantization is lossy: we will tell you the method and report the evaluation results honestly, and clause 19 applies to what we deliver.

19. Warranties and disclaimers

The Service is provided as available. Terms implied by statute that cannot be excluded by agreement (including, where it applies, the implied term that a service is performed with reasonable care and skill) are not excluded, and your remedies for their breach are the ones the law gives you. Beyond those implied terms we make no express warranty, and we offer no availability commitment: clause 7 says what that means in practice.

To the extent the law allows, Output is provided as is. Language models produce probabilistic text that can be wrong, incomplete, biased or offensive. We do not warrant that Output is accurate, complete, fit for any particular purpose, or free of third-party rights. Output is not professional advice. You are responsible for reviewing Output before you rely on it or pass it on, and that responsibility is heaviest where the consequences are.

20. Indemnity

You will defend us against third-party claims arising from Customer Content, from your use of the Service in breach of these documents or the law, or from your breach of an upstream model licence, and you will pay the damages and costs finally awarded or agreed in settlement.

We will tell you promptly in writing about a claim, give you control of its defence with counsel we can reasonably accept, and give you reasonable cooperation at your cost. We may take part with our own counsel at our own cost. You will not settle in a way that admits our fault, imposes any obligation on us or fails to release us fully, without our written consent, which we will not unreasonably withhold. If we materially prejudice the defence by failing to notify or cooperate, your obligation is reduced to the extent of that prejudice.

We do not offer an intellectual property indemnity for Output. We serve open-weight models under their upstream licences, we assign you whatever rights we hold in Output, and we do not take on the risk that a model’s training data infringes someone’s rights. If that is a requirement for you, tell us before you build.

21. Limitation of liability

Nothing in these documents excludes or limits liability for death or personal injury caused by negligence, for fraud or fraudulent misrepresentation, or for anything else that cannot lawfully be excluded. Subject to that:

Neither of us is liable for indirect, incidental, special or consequential loss, or for loss of profits, revenue, goodwill, anticipated savings, or loss or corruption of data.

Our total aggregate liability arising out of or relating to the Service is limited to the greater of (a) the amounts you paid us for Credits in the 12 months before the event giving rise to the claim, and (b) GBP 5,000. “Amounts you paid us” means what you actually paid for Credits, not what you consumed, and does not include Credits we granted. This is a single aggregate cap across all claims, not a cap per claim.

The cap does not apply to sums you owe us under clause 20, to either party’s breach of clause 13, to your obligation to pay for the Service, or to sums either party becomes liable to pay to a data subject or to a regulator as a direct result of the other party’s breach of the Data Processing Addendum. This clause is the single statement of how liability between us is limited; the Addendum’s liability clause points here rather than saying it again.

Nothing in these documents limits either party’s own liability to a regulator or to a data subject under data protection law: no agreement between us can cap what a regulator can fine either of us. Whether one party can recover its own regulatory fine from the other is a question the law answers, not this clause; the carve-out above preserves that claim between us without promising how the law resolves it.

22. Term, suspension and termination

This agreement runs until either of us ends it.

You may stop using the Service and close your Account at any time (clause 4).

We may suspend or terminate your access for material breach of these documents, or where clause 16 or the Acceptable Use Policy requires it. Where the breach can be fixed, we will normally tell you and give you a reasonable chance to fix it first. We may suspend immediately and without notice where the breach is serious or unlawful, where an authority directs it, or where continuing would cause imminent harm.

We may discontinue the Service on reasonable notice, returning purchased Credits you have not consumed.

Clauses 2, 11, 12, 13, 15 (as to amounts accrued, taxes and the refund position), 16, 19, 20, 21 and 24 survive termination, together with the Data Processing Addendum so far as it governs data we still hold, and any provision that by its nature should survive.

23. Changes

We may update these terms and any document incorporated into them (clause 3). This clause is the only mechanism for doing so.

For a material change we publish the new version with a new version number and effective date, and give reasonable notice before it takes effect, writing to Account holders at their registered email address. Non-material changes take effect on publication. From the first revision onwards, the outgoing version of a document stays available at a versioned URL so you can see what governed a past request.

If you object to a material change, tell us before it takes effect and you may terminate; we will return the purchased Credits you have not consumed. Using the Service after the effective date means you accept the change.

Two exceptions, both narrow. We may tighten a prohibition in the Acceptable Use Policy immediately where a change in law or an imminent risk of harm requires it. And sub-processor changes run on their own notice and objection process in the Data Processing Addendum, published on the sub-processor register, rather than under this clause.

24. General

Entire agreement. These terms and the documents incorporated into them (clause 3) are the whole agreement between us and replace anything said before. Neither of us relies on any statement not written here, and the Privacy Notice and Trust Center, which sit outside the agreement, are not terms. Nothing in this clause limits liability for fraudulent misrepresentation.

Severability. If a provision is unenforceable, it is modified to the minimum extent needed to make it enforceable while keeping its commercial intent, or struck if it cannot be. The rest stands.

Force majeure. Neither of us is liable for a failure caused by something beyond our reasonable control, including the failure of an upstream model provider or cloud platform. This does not excuse paying money that is owed.

Notices. Notices to you go to your Account email address. Contracts, subpoenas and formal legal notices to us go to legal@cyan.kiwi, or to the registered office for anything requiring formal service. Email notices are treated as received the next working day.

Assignment. You may not assign or transfer this agreement without our written consent, not to be unreasonably withheld. We may assign it to an affiliate or in connection with a sale of the business, on notice to you.

No waiver. Not enforcing something is not a waiver of it.

Third parties. Nobody other than you and us has any right to enforce these documents under the Contracts (Rights of Third Parties) Act 1999.

Governing law. English law governs these documents and any dispute arising from them, and the courts of England and Wales have exclusive jurisdiction.

If you are a consumer. The Service is not offered to consumers, and clause 4 asks you to confirm you are acting in the course of a business. If despite that you are treated as a consumer under law that cannot be excluded by agreement, nothing in these documents takes away the rights or the choice of forum that law gives you, and any provision here applies only so far as it lawfully can.

Publicity. Neither of us may use the other’s name or marks publicly without consent.