Choose a confidential GPU

Job type

Run inferenceInteractive APIs, agents, RAG and copilotsBatch inferenceDocument, media or dataset processingFine-tune a modelLoRA or QLoRA adaptationTrain a modelFull fine-tuning or continued pretraining

The estimate uses resident parameter count. For MoE models, enter total resident parameters.

Weight precision

BF162 bytes / weightFP8≈1 byte / weightINT8≈1 byte / weight4-bit≈0.5 bytes / weight