The short version: Use it lawfully, do not try to extract the model or the engine, and understand that model output can be wrong. We keep no record of what you send. Capacity is elastic, so treat availability as best-effort unless we have agreed otherwise in writing.
Acceptance
By calling the Sparkle API, directly or through a marketplace, you agree to these Terms. If you are agreeing for an organisation, you confirm you may bind it.
The service
Sparkle is an API that runs open-weight language models on our own inference engine and hardware, and
returns their output. It accepts text and, on models that support it, images and video, and returns
text. What each model accepts, its context length and its price are published by our
/v1/models endpoint, which is the authoritative description of the service at any moment.
The model is not ours
The model weights we serve are third-party open-weight models, published by their own authors under their own licences. Your use of the output is subject to the licence of the model you called as well as to these Terms. We do not extend any licence to the weights themselves.
Acceptable use
You may not use Sparkle to:
- break the law, or produce material that is unlawful where you or your users are;
- generate content sexualising minors, or content intended to harass, defraud or impersonate;
- attempt to extract, reconstruct or distil the model weights, or to reverse engineer the engine;
- circumvent rate limits, quotas, authentication or billing, including by distributing your key;
- send personal data you have no lawful basis to process, or special-category data you have not secured a basis for.
You are responsible for what your users send through your key and for what you do with the output.
Availability
Sparkle runs on elastic serverless capacity. Workers start on demand, so a request arriving at an idle endpoint waits for a worker to boot. We do not offer a uptime or latency commitment by default; where you need one, we will agree it separately in writing. We may change capacity, suspend a region for maintenance, or withdraw a model on notice.
Rate limits and suspension
We publish per-minute and concurrent limits and may enforce them by rejecting requests. We may suspend a key immediately, without notice, where we reasonably believe it is being used to break these Terms, to attack the service, or where a payment obligation is unmet.
Fees
Where you buy through a marketplace, that marketplace bills you on the prices we publish and its own terms govern payment. Where you contract with us directly, fees are as quoted, billed on metered token counts, and payable on the agreed terms. Prices may change with notice; a change never applies to requests already served.
Your content and our confidentiality
You keep all rights in what you send. We claim no ownership of prompts or of output. We do not retain, train on, or human-review your requests — see the Sparkle Privacy & Data Policy, which forms part of these Terms.
Output
Model output is generated statistically and can be wrong, biased, or fabricated. It is not advice of any kind. Identical prompts may produce output identical to that produced for another customer, so no exclusivity attaches to it. You must review output before relying on it, and you must not present it as human-authored where that would mislead.
Warranties and liability
The service is provided as is, without warranties of merchantability, fitness for a particular purpose, accuracy or non-infringement, to the fullest extent the law allows. We are not liable for indirect, incidental or consequential loss, nor for lost profits, revenue or data. Our aggregate liability for any claim is limited to the fees you paid us for the service in the three months before the claim arose. Nothing here excludes liability that cannot lawfully be excluded.
Indemnity
You will indemnify us against claims arising from your use of the service in breach of these Terms or of applicable law, including claims brought by your own users.
Termination
You may stop using the service at any time. We may terminate on notice, or immediately for a material breach. Because we retain no request content, termination leaves nothing of yours with us to return or delete beyond billing metadata.
Governing law
These Terms are governed by the laws of the jurisdiction in which InventDB Software LLP is incorporated, and disputes shall be resolved in the courts of that jurisdiction.
Changes
We may update these Terms; the version and date above will change. Continued use after a change is acceptance of it.