RunPod has shipped model caching that prevents every new GPU worker from downloading the same weights from Hugging Face or object storage. Model Store checks a host’s local disk first, then a datacenter-scoped network volume, and contacts the remote origin only when neither nearby tier has the files.
The scheduler now considers where a model physically resides before placing a worker. If a datacenter lacks a shared copy, one host downloads and verifies it while other workers wait; later workers read that shared cache rather than creating a transfer storm. Hashes identify model versions, and explicit states track downloading, verification, completion and failure. A stalled host is treated as such after about five minutes without progress.
The system falls back to the original source if the network volume fails or is full, so workers can still start without the cache benefit. Cached models are already available on RunPod Serverless, where each endpoint currently supports one cached model. Repositories with several quantization variants download all of them for now; selective downloads are only planned. A private repository with version pinning and direct export from a running pod remains in beta, so some of the broader version-management design is not generally available yet.