Fireworks AI officially releases the Training API and Fireworks Lab

On August 31, 2026, Fireworks AI announced the general availability of the Training API and Fireworks Lab. According to its introduction, the Training API allows users to write custom training loops in Python, with Fireworks hosting the distributed training and rollout infrastructure and handling gradient updates, sampling deployment, and weight synchronization. The API offers two compute options: Serverless mode bills by token and supports LoRA training; Dedicated mode bills by GPU hour, targets full-parameter training and larger context lengths, and requires consultation through Fireworks Lab to enable. Fireworks also claims it is one of the few organizations outside frontier labs running reinforcement learning on more than 10,000 GPUs. Official materials list usage results from teams such as Harvey, Vercel, Heidi, and Factory; for example, after Vercel fine-tuned an open-source model for v0's automatic fix feature, the error rate was 93% and end-to-end latency improved 40x. There are currently no subsequent development facts.

AI-generated from public reports

1. Aug 31· 2 reports

   Fireworks training API is officially available

   Source:Fireworks AI