Specialized Model Inference
Load an adapter
Load a LoRA adapter onto serverless with Specialized Model Inference.
POST
Load an adapter
Preview
Loads a LoRA adapter onto the shared serverless pool. Loading takes about 2 to 5 minutes. You’re billed per token only when you call the adapter.
The firectl equivalent:
Specialized Model Inference is in Preview and available on request. To request access, contact your account team or support@fireworks.ai. You can try the public adapter without access.
Path parameters
Body
The LoRA model to load:
accounts/{account_id}/models/{adapter_id}. It must be in the READY state, trained on a supported base model, with rank 64 or lower.Load onto one of your dedicated deployments instead:
accounts/{account_id}/deployments/{deployment_id}. The deployment must have addons enabled. Billed by GPU-hour, not per token.Response
Returns the deployed model. It starts inDEPLOYING; send traffic once it’s DEPLOYED. Save name: its last segment is the deployed model ID you need to unload over REST.
Behavior
- One serverless copy per adapter. Loading the same adapter onto serverless again fails while the first copy exists.
- The first request is slower. The first request after loading a LoRA adapter can take up to a minute. Later requests have much lower latency.
- The bare model name. The first copy of an adapter you load becomes its default and answers
accounts/<ACCOUNT_ID>/models/<LORA_ID>. If another copy already holds that name, this copy loads withdefault: false. Call it withaccounts/<ACCOUNT_ID>/deployedModels/<DEPLOYED_MODEL_ID>. There is no#serverlesssuffix. See Which model string to use. - Adapter limit. You can load up to 3 adapters per base model on serverless at a time. Going over it returns
429 RESOURCE_EXHAUSTED. Unload an adapter to free up a slot. For more slots, contact support@fireworks.ai or your account team. See Supported models and limits.
Errors
Load onto a dedicated deployment instead
Load an adapter