Model status
The status page shows how each model in the Axelera AI lineup is behaving right now, so you can tell a bad answer apart from a bad day.
The status page lives at /status. It shows one row per model in the lineup — axelera-mercury-1.0, axelera-venus-1.0 and axelera-earth-1.0 — each with a single label. These are the same names you send as model when you call the API, so what you read here is what you are calling.
There is no menu entry for it, by design: it is a page you bookmark. Whenever a warning strip appears at the top of the app, View status on that strip takes you straight here.
Status is measured over a rolling fifteen-minute window and refreshes about once a minute. It describes the recent past, not this second.
| Label | What it means | What to do |
|---|---|---|
| Operational | Almost every recent request on this model came back with an answer. | Nothing. This is the normal state. |
| Degraded | A noticeable share of recent requests failed or came back empty. | Answers may be slower. Chat uses whatever is healthy behind the model automatically. |
| Down | Most recent requests on this model failed or came back empty. | Chat uses another model automatically. If answers are still failing, try again in a few minutes. |
| No recent data | Too few recent requests to judge. Usually a quiet lane, not a problem. | Nothing. A model can sit here for hours on a healthy day. |
What counts as a failure
What a model means here
A model in the lineup is a promise about the answer you get, not a single engine. Behind each one Axelera picks and re-picks what actually runs.
That is what Degraded usually means here: the model is still answering, but not everything behind it is healthy, so replies can be slower and there is less to fall back on. Down means nothing behind it is answering.
A banner appears at the top of the app only when nothing healthy is available across two consecutive windows, which is the point at which answers really are affected. One bad window never raises it, and one clean window clears it.
What the page does not tell you
The page reports whether a model is answering, and nothing about why. It does not name a provider as the cause, because a failed request cannot always be told apart from a limit reached on our side, and blaming the wrong party is worse than staying quiet.
Speed is also out of scope. A slow model is not a failing one, so a model answering slowly still reads as Operational.