Hugging Face
Hugging Face Changelog
The Hub is building out infrastructure for agents as first-class entities, shipping storage and filesystem layers alongside trace capture and direct agent integration. hf-mount graduated remote storage to native filesystem access, letting agents treat Hub buckets as local disks 100x larger than physical hardware. Spaces now mount buckets directly as volumes, and the new Agent Traces viewer auto-detects and organizes debugging sessions from Claude Code and similar tools. The Hub also shipped protected Spaces—keeping model weights private while serving the app publicly—and wired Hugging Face Papers into agent tooling so Claude and Cursor can search and read papers without token bloat.
Storage and access patterns dominated the month. Spaces gained persistent storage through HF Buckets with direct mounting, while hf-mount enabled attaching remote storage 100x larger than local disk as a filesystem—crucial for agentic workloads handling massive model weights. Protected Spaces let developers keep URLs public while hiding implementation details, and AI agents gained native Hugging Face Papers integration with semantic search across papers and linked resources.