Skip to main content
Start with current offers; a model appearing in documentation does not guarantee it is in stock.

​
Check these fields

  • VRAM per GPU: allow space for model weights, activations, context, and your application’s overhead.
  • GPU quantity: multiple GPUs require software that can use them. Their memory is not automatically one shared pool.
  • Whole-instance hourly price: compare the configuration, not just the model name.
  • Maximum duration and budget: make sure the available runtime covers your job.
  • Storage: copy important results off the instance before it ends.
A 24 GB card may suit smaller models and image workflows. Larger-memory cards may be needed for larger models or batches. Fit and performance depend on precision, architecture, context length, and software; Ukiyo does not promise a particular model will fit from its parameter count alone. Use a small representative workload before committing to a long run. See SSH and Jupyter for access; application installation is separate from renting compute.