Model downloads
Local providers can fetch models from two places: the provider’s own registry, and Hugging Face. Both are reached from the download dialog in Local Models. Which sources a provider offers is listed in Providers; pasting a Hugging Face URL switches the source for you.
Registry or Hugging Face
Section titled “Registry or Hugging Face”The registry expects a name exactly as the provider lists it, tag included. It is the shorter
path when the provider publishes a curated build of the model you want. Hugging Face takes
either a model card URL or a publisher/model-name id, and the repositoryis read before
anything downloads. Cannonade reads it anonymously, so a private repository cannot be resolved
at all.
Quantization
Section titled “Quantization”A Hugging Face repository usually ships the same model at several quantization levels.
Cannonade preselects Q4_K_M when the repository has it, otherwise takes the first available.
The choice matters more than it looks: it decides how much memory the model needs and how much quality it loses. Picking a level your machine cannot hold is the common cause of a model that downloads fine and then fails to load.
Repositories with no GGUF quantizations offer no picker, and the provider takes whatever it can use.
Registry downloads have no picker either. The tag in the name already identifies the build, so quantization is part of what you typed.
Downloading for a run
Section titled “Downloading for a run”A test run does not have to be limited to models you already have. Naming a model that is not installed makes the run download it before the cases start.
Delete auto-downloaded models after the run cleans those up again. It only removes models that run actually downloaded: a model that was already installed is left alone, even when the same run used it.
Downloads are per model run, so a failed download fails that model and leaves the rest of the run going.