Post-Training pricing

Training is priced per GPU-hour and prorated by active training time. Each GPU comes pre-provisioned with the base model weights and training image, so you pay only for active training time. Inference, evaluation, and checkpoint storage are billed separately.

Free during Private Preview

Model Distillation uses Serverless SFT to train a smaller, faster model to reproduce a larger model’s behavior on a specific task. See the documentation to learn when to use Model Distillation.

Model Distillation is not billed separately, but uses hosted CoreWeave services to collect traces, generate datasets, train models, and serve them for inference.¹