Send a test request
Datacenters
Recent Routes
| Source | Destination | Model | Status | Distance | Why |
|---|
Models across the fleet
| Model | Type | Available on | Warm on |
|---|
How routing decisions are made
Stay where the model is already warm
If a model is already loaded somewhere in the fleet, new requests for it are biased toward that region instead of triggering a redundant cold load elsewhere -- better latency, less duplicate GPU memory spent on the same weights.
Nearest capable node, unless it's busy
Distance to the client is weighed against each node's current load. A node at capacity is excluded entirely, so traffic gracefully overflows to the next-best region instead of queuing.
Don't use the big GPU for a small job
Every node advertises a cost tier (premium GPU, standard GPU, or CPU). When a model is warm on both an expensive and a cheap node, the cheap one wins unless it's saturated or too far away to make sense.
Conversations stick to their node
Each chat session carries an id. The node that served the last turn keeps serving it -- losing that would mean reprocessing the whole conversation, not just a cold model load, so this bonus outweighs generic model affinity.