Fireworks opens training API after 10,000-GPU RL

Fireworks made its Training API and Fireworks Lab generally available and says it has operated reinforcement learning across more than 10,000 GPUs, according to the launch post. Its serverless path trains LoRA adapters with per-token billing, while dedicated runs support full-parameter training and charge by GPU-hour; both can deploy a checkpoint to inference without separate rollout infrastructure. Teams evaluating tuning should compare total training-plus-serving cost, not a headline API rate—recent model price-per-point analysis makes the same accounting case.