WRITE LOCAL CODE · USE CLOUD GPUS
Point Fleet at a model and a dataset and it provisions the GPU, runs the job, and streams loss and learning rate live. LoRA, QLoRA, full, and DPO, all from plain Python.
Full fine-tuning, LoRA, 4-bit QLoRA, and DPO. Switch with one hyperparameter; sensible defaults are pre-filled.
Fleet estimates the VRAM your run needs and picks the cheapest GPU that fits. Pin a tier yourself if you prefer.
Loss, learning rate, and validation streamed every few steps and stored server-side, so you can watch a run or review it later.
Save and resume from intermediate checkpoints. Fleet prunes old ones automatically to keep storage costs down.
Pay only for the GPU time the job actually uses. Cancel early and you stop paying immediately.
When the job finishes, the fine-tuned model is ready to deploy or keep training from.