Hugging Face Hub rearchitects uploads/downloads using CAS; regional PoPs planned in us-east-1, eu-west-3, ap-southeast-1

AI Impact Summary

Hugging Face Hub is shifting to a content-addressed store (CAS) as the first hop for content distribution, with reads routed through CAS and writes performed on chunks backed by an S3 store in us-east-1. The redesign targets large model files (e.g., 131GB weights across 30 files) and aims to enable byte-level deduplication, compression opportunities (e.g., Safetensors), and tighter data integrity via an operational control plane and enhanced telemetry. Regional CAS points of presence are planned (us-east-1, eu-west-3, ap-southeast-1) with a phased production by end of 2024, and an expected bandwidth efficiency (~12% lower total upload bandwidth) despite increased first-hop distance. This promises improved scalability for large artifacts and enterprise use cases, but will require client tooling updates to adopt the new protocol and paths.