Skip to content

Ask AI

I answer questions about Radeon Cloud from these docs and link to the pages I used.

Answers are generated from the docs by a model on Token Factory and can be wrong; the docs are authoritative. Don't enter API keys or other secrets here.

Model load

Returns the current load of each Token Factory model, as reported by the inference engines serving it. Use it alongside List models: pick a model from the list, then check here how busy it is.

GET /v1/models/utilization API key or session
GET /v1/models/utilization/{model} API key or session

Also available as /api/v1/models/utilization.

No parameters.

Terminal window
curl https://developer.amd.com.cn/radeon/api/v1/models/utilization \
-H "Authorization: Bearer $RADEON_API_KEY"
{
"object": "list",
"data": [
{
"id": "DeepSeek-V4-Flash",
"utilization": 16.3,
"state": "idle",
"partial": false,
"sampled_at": "2026-11-01T00:20:15.458Z"
},
{
"id": "MinerU2.5-Pro",
"utilization": null,
"state": "unknown",
"partial": true,
"sampled_at": null
}
]
}

Put the model name in the path. Matching is case-insensitive.

Terminal window
curl https://developer.amd.com.cn/radeon/api/v1/models/utilization/DeepSeek-V4-Flash \
-H "Authorization: Bearer $RADEON_API_KEY"
{
"id": "DeepSeek-V4-Flash",
"utilization": 16.3,
"state": "idle",
"partial": false,
"sampled_at": "2026-11-01T00:20:15.458Z"
}

A model that isn’t served returns 404:

{
"error": {
"message": "Model GPT-5 is not available",
"code": "model_not_found"
}
}
FieldDescription
idModel name, the same id as in List models.
utilization0 to 100, the share of the model’s serving capacity in use. null when no data could be collected.
stateidle (below 50), busy (50 to 90), full (90 or more), or unknown when there’s no data.
partialtrue when fewer instances were sampled than are running. If utilization has a value, it is estimated from the sampled ones.
sampled_atTime of the oldest sample behind the figure, ISO 8601.

Load is for display only and does not tell you whether a request will be rate limited. A model in full may still accept your request, and one in idle may still return 429 because of your own quota. See Rate limits.

A null utilization doesn’t mean idle. It means there’s no data right now; treat it as unknown. If utilization is null, partial: true means the model has running instances that couldn’t be sampled, and partial: false means none are running.

The OCR and speech models (MinerU2.5-Pro, OmniVoice) and Clef don’t report load, so they currently always show unknown with partial: true.

Data refreshes about every 10 seconds and is cached for 5, so polling faster gets you nothing new. The endpoint doesn’t count against your concurrency allowance, but it does count against the per-minute limit.

Returns 503 when the data is temporarily unavailable:

{
"error": {
"message": "Model utilization is temporarily unavailable"
}
}

Unlike the 404, this body has no code field.

The load bar on the model cards in Token Factory comes from a different counter, so the two can show different numbers.