Settings

Which model

The list comes from this repository's models.json. “custom” accepts another hub repository. “on this device” reads a file or folder from disk — nothing is copied or cached.

no model loaded

Limits applied the next time a model loads

Context is conversation plus reply held at once. Every token of it is scanned on every step, so a larger context is a slower reply — left empty it picks a size that stays responsive. GPU memory is guessed from system RAM: right for unified memory, too high for a discrete card — set it to your VRAM if you have one.

Environment
detecting environment…