What changed
Hugging Face has released new features for its CLI and huggingface_hub library, including shared caching, improved Jobs, faster Storage Buckets, and enhanced DX for AI agents. The shared cache feature ensures that files are downloaded only once across repositories, reducing disk usage and transfer time. Hugging Face Jobs now support retries, reruns, and UV script configurations, with better behavior in scripts and agents. Storage Buckets have faster downloads and live listings, and Sandboxes allow running code in isolated cloud machines. AI agents are now a significant part of hf CLI users, with recent models already familiar with hf CLI commands.
What this means in practice
The shared cache feature reduces redundant downloads and disk space usage by linking repositories to a single cached file. Users can manage cached files individually or prune unused ones. Hugging Face Jobs now offer more flexibility with retries, reruns, and scheduling, and can be triggered by webhooks. Storage Buckets benefit from faster downloads using hf_xet and live file listings, improving data access. Sandboxes provide isolated environments for running code, useful for agents needing to execute code without local dependencies. AI agents are now a major user group, with the hf CLI skill helping them perform tasks more efficiently.
Limitations and availability
The shared cache feature is available without configuration, and users can manage cached files with commands like hf cache rm and hf cache prune. Hugging Face Jobs improvements are part of the latest release, with some features like UV script configurations and webhook triggers now available. Storage Buckets' faster downloads and live listings are enabled through the HfFileSystem library, which is part of the huggingface_hub library. Sandboxes are experimental and run on Jobs, requiring Docker images. The hf CLI skill for AI agents is available in recent models like DeepSeek-V4.1-Flash and Kimi-K3, though it does not improve accuracy but reduces the number of steps needed to complete tasks.
