Fireworks AI Launches Training API for Custom ML Loops

On August 31, 2026, Fireworks AI announced the general availability of its Training API and Fireworks Lab. [1] This API connects a customer's Python training loop to distributed compute managed by Fireworks. [1] Under this arrangement, customers orchestrate the loop and maintain control over the data, environment, and loss or reward function. [1]

Meanwhile, Fireworks manages the interaction between the trainer and rollout, which includes weight synchronization and train-rollout alignment. [1] The company stated this release is a response to ML team constraints like fragmented training infrastructure and restricted parameter control. [1]

The Training API provides two distinct compute options: serverless and dedicated. [1] The serverless tier is billed per-token and allows customers to train LoRA adapters on an always-on shared pool with a curated list of models. [1] In contrast, the dedicated tier bills per GPU hour and supports full-parameter training for the largest mixture-of-experts models without contention or rate limits. [1] Once trained, promising checkpoints can be deployed to production inference through the UI or API. [1]

By uncoupling the orchestration of the training loop from the underlying compute infrastructure, Fireworks AI is attempting to bridge the gap between fully managed, rigid platforms and highly complex, self-managed clusters. The dual-tier approach offers a path from shared-infrastructure LoRA experimentation to full-parameter dedicated tuning without changing the underlying orchestration layer. What the sources don't address: Pricing specifics for both the per-token serverless tier and the per-GPU-hour dedicated tier are not detailed.