A new managed option now covers custom training loops on open models. Fireworks AI moved its Training API to general availability on August 31, bundling it with Fireworks Lab for rollout infrastructure.
The customer’s own Python loop stays in charge, with full control of the loss or reward function, data, and environment. Weight synchronization, failed-swap recovery, and trainer-rollout alignment all sit with Fireworks, which runs the distributed compute layer beneath the loop.
Two compute tiers come with the API. Serverless training runs LoRA adapters on a shared pool billed per token, aimed at experiments and reinforcement learning loops. Dedicated training handles full-parameter runs on the largest mixture-of-experts models, billed per GPU hour on elastic capacity.
Fireworks frames the launch as an answer to complaints about existing workflows: limited model choice, tight control over training loops, rigid compute, and training and rollout systems that don’t fit together.
