Training
Adapter merging
What is adapter merging?
Adapter merging folds a trained adapter’s weights into the base model to produce a single standalone model. It removes the small inference overhead of applying the adapter separately, at the cost of losing the ability to swap or share it.
A merged model is easier to hand to someone who does not know or care about adapters, and it is the form most third-party tooling expects. It is also the form to publish when the goal is for others to use the model directly.
The trade is co-residency. Once merged, the model no longer shares a base with its siblings, so serving ten merged variants costs ten times the memory of serving ten adapters against one base.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.