Logical models

A logical model is the model name a client writes in its request, and it is where routing lands. Underneath it hangs a set of real provider models; once a request lands on it, one of them is chosen to actually run.

The Logical models page in the left sidebar manages request priority, switching mode, and model enablement.

One card, one logical model

Each logical model is a card:

Drag to reorder to set priority between logical models.
The model rows inside a card can be dragged too, deciding the try order during failover.
Add model creates a scheduling entry for the current logical model.

No explicit add, no participation

Only models explicitly added to a logical model take part in requests. Adding one creates a scheduling entry.

Two modes

Each logical model card has a toggle:

Mode	Behavior
Failover	Tries the attached models in order, automatically switching to the next on failure.
Manual pinning	Requests use one fixed upstream model (that row is marked "pinned", the rest stand by).

Add a model before switching to manual

You cannot switch to manual pinning mode without any models.

Metric cards

The top of each card shows four metrics to help you judge who should rank first:

Metric	Meaning
Success rate	Success rate over the last N completed requests.
Average latency	Full latency of the attempt that served the request, including TTFT.
Average TPS	Output tokens ÷ full latency of that attempt (including TTFT).
Currently available models	Number of models available right now, and whether failover happened recently.

Each model row also shows: the time of its last success, its consecutive failure count, whether it is cooling down, and whether it is disabled.

The built-in `default`

default is the fallback logical model: any request that matches no other logical model lands here.

Its ID cannot be changed; its description still can.
It cannot be deleted.
It is the final safety net for every strategy in Request routing.

Create & edit

When you create a logical model, fill in:

Logical model ID: starts with a letter or digit and may contain letters, digits, dots, underscores, and hyphens, up to 64 characters. The model name in a request must be usable verbatim as this ID, so dots in version numbers are allowed (e.g. deepseek-v4.1-flash).
Description (optional).

The ID is the identity

Requests send it as the model name and the routing graph references it by it. Renaming swaps all of those references in the same save.

Delete a logical model

After deletion the UI no longer shows it, and requests that matched it fall to default.

Once logical models are arranged, go see how requests are actually routed.