The article addresses the "broken" scale of AI model distribution, where transferring massive models (like a 130GB DeepSeek-V3) to hundreds of GPU nodes simultaneously bottlenecks network bandwidth and hits rate limits on hubs like Hugging Face. The solution is the CNCF project Dragonfly, which uses a Peer-to-Peer (P2P) architecture to ensure a model is fetched from the origin only once. It details new native protocol support for hf:// and modelscope://, allowing dfget (Dragonfly's download tool) to natively understand these hubs, reducing origin traffic by up to 99.5% and enabling wire-speed distribution across clusters.

https://www.cncf.io/blog/2026/04/06/peer-to-peer-acceleration-for-ai-model-distribution-with-dragonfly/