Model APIs

AMD Radeon Cloud serves models behind an OpenAI-compatible HTTP API. Any client that talks to OpenAI works by changing the base URL and the key: curl, the openai SDK, LangChain, Cherry Studio.

Two options, and the difference is who owns the GPU.

Free shared endpoints

Shared endpoints are always on and cost nothing. No instance to launch, no credits spent. The platform picks which models are available.

Open the Token Factory and sign in.

Under Public Free Model APIs, pick a model. The detail dialog shows the base URL, the model name, your API key, and a ready-to-run curl command.

curl https://developer.amd.com.cn/radeon/api/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"DeepSeek-V4-Flash","messages":[{"role":"user","content":"Hello"}]}'

One key covers every shared model. To use a different one, change the model field.

Shared endpoints support chat completions and listing models. Requests are rate limited per key and per IP, and a daily spend cap applies — see Rate limits.

Dedicated endpoints

A dedicated endpoint runs the model on your own instance, so you choose the model and the serving configuration. It uses credits for as long as the instance is up.

1. Go to Profile → Add Template in the console.
2. Set Deploy Type to vLLM Model API, then write the serve command.

vllm serve Qwen/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 8000

Keep --host 0.0.0.0 --port 8000

The platform routes the public endpoint to port 8000 inside the container. If the server binds anywhere else, the endpoint won’t reach it.

3. Save, then Launch the template. When the instance is ready, the dialog shows a Base URL of the form https://<host>/spaces/<instance-id>/8000/v1, plus the model name and API key.

Call it exactly like a shared endpoint, with that base URL:

curl https://<host>/spaces/<instance-id>/8000/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen/Qwen2.5-7B-Instruct","messages":[{"role":"user","content":"Hello"}]}'

Because the request goes straight to vLLM or SGLang, a dedicated endpoint exposes whatever that server exposes — typically completions and embeddings in addition to chat.

Choosing

Start with shared endpoints. Move to a dedicated one when you need a model that isn’t in the catalog, want control over serving parameters, or need throughput that isn’t shared with other users.

Full request and response details are in the API reference.