API overview

The AMD Radeon Cloud API has four parts:

• The Token Factory API calls the models hosted in Token Factory. It’s OpenAI-compatible, with an Anthropic-compatible Messages endpoint.
• Model service endpoints reach vLLM running on your own instance, launched with Deploy type: Model API. They’re also OpenAI-compatible, but run on a different host with different limits and error bodies.
• CLI endpoints are what Radeon Cloud CLI calls. Use the CLI to script instances.
• The Platform API serves the web console: instances, templates and your account. It accepts only a browser session.

Base URLs

API	Base URL
Token Factory API	https://developer.amd.com.cn/radeon/api/v1 (China site only)
Model service endpoints	Shown on the instance card, for example https://jupyter.anruicloud.com/inference/<instance-id>/v1
Platform API	https://developer.amd.com.cn/radeon or https://radeon-global.anruicloud.com

Token Factory is available on the China site only.

A model service’s base URL and key are shown on the instance card once it’s ready. Copy them from there instead of building the URL yourself.

Prefer what the console shows

These are the public addresses at the time of writing. If the console or Token Factory shows a different base URL, use that one.

Credentials

Token Factory takes your rc- API key, a model service takes the sk- key on its card, the CLI uses its own rcc_ token, and the Platform API takes a browser session. See Authentication.

Conventions

Requests and responses are JSON with Content-Type: application/json. Timestamps are ISO 8601. Errors have a non-2xx status and a JSON body. See Errors.

Token Factory API

Always on, nothing to launch, and no credits spent. AMD chooses which models are served. The main operations:

• POST /v1/chat/completions: OpenAI-compatible chat
• GET /v1/models: the models you can call
• POST /v1/messages: Anthropic-compatible, for clients such as Claude Code
• POST /v1/messages/count_tokens: the token count those clients ask for first
• GET /v1/usage: your usage and remaining daily allowance

Some models use other paths under /v1, such as /v1/audio/speech for speech or /v1/ocr for document parsing. Each model’s page in the model catalog lists the paths it uses. A path that no model serves returns 404.

Every path works with and without the /api prefix: /api/v1/models and /v1/models are the same endpoint.

Requests go through a gateway that checks them against a fixed set of fields, so a parameter outside that set is dropped rather than passed on. See Chat completions for the list. Per-key rate limits, a concurrency allowance and a daily usage cap apply. See Rate limits.

Model service endpoints

vLLM running on an instance you launch with Deploy type: Model API. You choose the model and the serving flags, and the instance spends credits while it runs.

Requests are passed to your server unchanged, apart from the routing prefix. Whatever vLLM serves is available, usually including /v1/completions and /v1/embeddings. There’s no gateway in front, so Token Factory’s rate limits, model list and error bodies don’t apply.

See Model service endpoints.

Platform API

Area	Reference
Launch, check and destroy instances	Instances
Gallery and model service templates	Templates
Profile, resources, SSH keys, coupons	Account

Next steps

Read Authentication next. Sending the wrong credential is the most common cause of errors.