A decentralized network Pirate Face has been launched for mirroring AI models

Developers presented Pirate Face — a decentralized P2P infrastructure designed for mirroring open AI models from the Hugging Face platform. The project ensures resistance to censorship and guarantees constant access to model weights through a torrent distribution system. Each file in the network undergoes a checksum verification procedure, which eliminates the possibility of data substitution or damage to archives during transmission. The tool allows users to download models directly from other network participants, bypassing centralized servers that may restrict access based on geographical or legal grounds. The project's architecture is focused on creating a "kill-proof" archive of weights that maintains functionality even when main repositories are disconnected. The use of the BitTorrent protocol provides high download speed due to the distribution of load between network nodes, which is critical for models with volumes in tens and hundreds of gigabytes. The solution is aimed at preserving open developments in the face of tightening content moderation policies on major hosting sites. The project has already received attention from the community on Hacker News, where issues of long-term data storage and accessibility for researchers are discussed. The launch of Pirate Face changes the approach to weight distribution, moving them from centralized assets to distributed data, reducing the risk of sudden removal of models from public access and ensuring the independence of the infrastructure from the decisions of individual corporations.