Skip to main content
Impala serves models fine-tuned with LoRA (low-rank adaptation). You bring the adapter, we deploy it with its base model, and you send requests to it through serverless or batch.

​
What to send us

Either form works:
  • A separate adapter, alongside the name of the base model it was trained on.
  • A merged model, with the adapter weights already folded into the base model.
Contact your Impala account team to share it.

​
How it’s served

The adapter is loaded with the model when it’s deployed, so every request to that deployment uses it. Switching adapters per request isn’t supported yet.

​
Sending requests

Send requests exactly as you would. See Chat completions for serverless and Run a batch for batch. If you retrain full weights often, Impala Leap swaps new checkpoints onto a running deployment without a redeploy.