API terms of use
Last updated: 2026-09-10
These terms cover the Model API — the free shared endpoints documented in this section. They do not govern GPU instances, templates, or the rest of the platform.
The Model API is offered by the AMD developer community as a trial and development service for people building on AMD hardware. It is free, and it is provided as is.
By requesting an API key and calling these endpoints, you accept what follows.
What it is for
Section titled “What it is for”Evaluate the models, prototype against them, and decide whether AMD accelerators fit your work. Typical uses are evaluation, experimentation, learning, demos, hackathons, and open-source development.
Not for production
Section titled “Not for production”Do not put a production workload behind these endpoints. There is no service level agreement, no uptime commitment, and no support obligation.
Concretely:
- Models can be added, changed, or withdrawn from the catalog with little or no notice.
- Endpoints and API keys can be restarted or revoked for operational reasons.
- Capacity is shared. Requests can be rate limited or rejected when the service is busy.
- Nothing here is designed for high availability.
If your work depends on the endpoint being up, run it somewhere with a commercial agreement behind it.
The models are the upstream originals
Section titled “The models are the upstream originals”Models are served from their published upstream weights, unmodified, running on AMD accelerators. We do not fine-tune them, distill them, or substitute a different model behind a familiar name.
Two consequences follow:
- The behaviour, capabilities and limits are those of the upstream model. Read its model card. Its licence applies to your use of its outputs.
- We do not warrant that any output is accurate, safe, current, or fit for any purpose. Review anything you rely on.
The model name a response reports is the model that produced it. If you reached this service through a third party who told you it was a different vendor’s model, you were misinformed — see Reselling.
Free of charge
Section titled “Free of charge”The Model API is free. Credits, quotas and rate limits are capacity controls, not currency, and carry no monetary value. Free also means no purchased entitlement: quotas can change and access can be suspended without compensation.
Keys and accounts
Section titled “Keys and accounts”One person, one API key. You are responsible for everything done with your key, so keep it secret and rotate it if it leaks.
The following will get a key disabled:
- Holding keys on more than one account for the same person, including through
address aliases such as dots or
+tags in an email address. - Sharing, transferring, or selling a key.
- Automating account creation to obtain more keys.
Reselling is not allowed
Section titled “Reselling is not allowed”You may not resell, sublicense, or otherwise provide access to these endpoints, whether for money or not. This includes:
- Exposing the API to third parties, directly or behind a proxy, gateway or aggregator.
- Presenting the service, or its models, as your own product or as another vendor’s product.
- Instructing a model to misreport which model or which provider is answering.
- Pooling accounts or keys to raise the capacity available to a group.
Anyone can register directly and use these endpoints for free at the Token Factory. There is no reason to pay a third party for access, and no third party is authorised to sell it.
Keys found to be reselling are disabled. Calls made with a disabled key are answered with a notice explaining why, so that anyone who obtained access through a reseller can see what happened and register directly.
Acceptable use
Section titled “Acceptable use”Do not use the Model API to:
- Break the law, or infringe anyone’s rights.
- Generate malware, exploits, or tooling for unauthorised access to systems.
- Attack, probe, or overload any system, including this one — testing your own systems is fine, testing other people’s is not.
- Circumvent quotas, rate limits, authentication, or content safeguards.
- Generate sexual content involving minors, or content that harasses, defames, or incites violence against people.
- Impersonate a person or an organisation, or generate content presented as authored by one.
- Process personal, health, financial, or otherwise sensitive data. This is a trial service; do not send it anything confidential.
What we log
Section titled “What we log”We log API requests, including the content of prompts and model outputs, together with metadata such as timestamps, model, token counts and errors.
- Conversation content is kept for a short rolling window — currently two days — and then removed. Metadata is kept longer for usage accounting.
- Content is used only to operate the service: capacity planning, debugging, abuse detection, and enforcement of these terms.
- We do not train models on your data, and we do not sell it or share it with third parties for their own use.
Content may be retained beyond the normal window when needed to investigate a specific abuse case or to comply with a legal obligation.
We are telling you this because it is true, not because it is comfortable: if you would not want a prompt stored, do not send it.
Suspension
Section titled “Suspension”Keys that break these terms may be disabled without notice. We may also disable a key when we reasonably suspect abuse and are still investigating.
If you think your key was disabled in error, write to member-service@mail.developer.amd.com.cn.
Availability and changes
Section titled “Availability and changes”We may change, suspend or discontinue any part of the Model API, and we may amend these terms, at any time. Material changes will be reflected here with a new date at the top. Continuing to call the endpoints after a change means you accept it.
No warranty, no liability
Section titled “No warranty, no liability”The Model API is provided as is and as available, without warranty of any kind. To the extent permitted by law, we are not liable for any loss arising from your use of it, including lost data, lost profits, or the consequences of acting on model output.