Skip to main content
The Training API uses the same Tinker-compatible primitives on two infrastructure paths. Choose the path before adapting a cookbook recipe.

Serverless Training

Attach to a shared pooled trainer. There is no trainer or rollout deployment to provision.

Dedicated Training

Provision trainer and deployment resources for your run, with broader model and method support.

Quick decision

Comparison

Always verify current models, limits, prices, and feature status in the Serverless Training and Dedicated Training pages before launch.

Choose serverless when

  • The base model appears in the current serverless model list.
  • LoRA SFT or RL covers the task.
  • You want to start without provisioning trainer or inference resources.
  • Per-token billing fits a small or bursty experiment.
  • In-session sampling is sufficient.

Choose dedicated when

  • You need full-parameter training, DPO, ORPO, distillation, or a model not on the serverless list.
  • You need explicit trainer, rollout deployment, checkpoint, reconnect, or promotion control.
  • You need sustained throughput or long-running rollouts.
  • A highly utilized time-based deployment is more economical for the workload.
  • You need to serve or evaluate through a dedicated deployment during training.

The interface is a separate choice

Serverless and dedicated describe how training compute is provided. They are not separate coding-agent modes. You can ask the Fireworks training skill to choose and run either path. You can also use the Python SDK or cookbook directly. Managed fine-tuning is a separate workflow for standard jobs where Fireworks owns the training loop.

Next steps