What is the difference between serverless and on-demand?

Serverless is pay-per-token inference on shared Fireworks capacity. On-demand is a deployment on GPUs reserved for your account.

Serverless is shared inference. You send a request to api.fireworks.ai and pay per token. You do not choose the GPU or the replica count. Use it for models Fireworks already hosts, and for LoRA adapters on select base models.

On-demand is a deployment on GPUs reserved for your account. You pay for GPU time, and you choose the model, hardware, and scaling. Use it for a custom model, a LoRA adapter on a base model Serverless does not support, or a latency target Serverless does not meet. Docs also call this a dedicated deployment.

The comparison table is on Serverless overview. To start, use the Serverless quickstart or the on-demand quickstart.