A new service called Pirate Face is offering a decentralized method to preserve open-source AI models hosted on Hugging Face, ensuring they remain accessible even if the original platform removes them. The system mirrors every model from Hugging Face as a peer-to-peer torrent, storing copies across a global swarm of users rather than relying on a single corporate server. Each file includes its official Hugging Face SHA-256 checksum, allowing users to verify that downloaded weights match the original, unaltered version. When a model is taken down from Hugging Face, the torrent continues to function through the web-seed mechanism, which initially pulls data directly from Hugging Face and later shifts to distribution via the peer network. This approach eliminates a single point of failure, as the model can be retrieved from any participant in the swarm. Pirate Face currently requires users to have a Hugging Face account to submit models, though browsing, downloading, and seeding do not necessitate an account. Verification of a creator’s identity is possible through matching Hugging Face account credentials, which helps prevent impersonation while maintaining independence from centralized hosting. The platform plans to introduce direct model publishing in the future, reducing reliance on external services. Downloads are facilitated through a drop-in API that redirects existing pipelines to the Pirate Face endpoint, requiring no code modifications. While the service draws on Hugging Face for initial distribution, its long-term objective is to sustain model availability beyond the lifespan of any single hosting provider, thereby supporting the continuity of open-source AI development.