Skip to main content
  • Planning estimates are not quotes.
  • Managed estimates can be low when rendered-token inputs omit multi-turn unrolling. Dedicated compute floors can be low when sequences pack poorly.
  • Dedicated shows a saturated compute floor. Allocated GPU time spent initializing the model, writing checkpoints, or idle can make real jobs cost more. Queue wait and pre-allocation provisioning are not billed.

​
Prepare inputs with the Training Skill

The Fireworks Training Skill prepares inputs and calculates Managed and Serverless estimates from published rates. It can inspect a local dataset or an existing job and identify assumptions. It does not calculate Dedicated numbers. Use this page for Dedicated planning. The Skill does not launch a job or authorize spend while estimating. Training still requires the Skill’s complete final plan and your explicit confirmation.

​
Reinforcement learning

RL cost varies with rollout shape, concurrency, reward or verifier design, training method, and evaluation workload. Contact the Training team for a tailored estimate. To compare the rollout inference cost of multi-turn agentic RL, use the separate rollout cost comparison. That page does not estimate SFT or DPO training cost.