Hugging Face Hub 2.2.0 changes how job metrics behave by default, and fixes download and filesystem issues that could give developers stale or incomplete results. The practical catch: scripts that relied on a live stream now need to opt in to following updates.
Hugging Face Watch analysis
What happened
The Hugging Face Hub 2.2.0 release notes, published on 8 October, describe a release of bug fixes and quality-of-life improvements. Most notably, hf jobs stats now prints one snapshot and exits instead of streaming until a job finishes. Use -f or –follow for live updates; the Python fetchjobmetrics function also defaults to a snapshot, with follow=True restoring streaming.
The release also fixes a race that could make concurrent downloads skip hfxet, and updates several HfFileSystem cache behaviours. Here are five useful changes in the notes:
Our top picks
- Job stats return promptly
The default snapshot avoids scripts or agents waiting on a stream that never ends. - Live monitoring remains available
Use -f in the CLI or follow=True in Python when you want continuing metrics. - Concurrent downloads consistently use hfxet
A version-lookup race could send some files through plain HTTP instead. - Filesystem listings refresh cleanly
Refreshed listings replace cached results, avoiding duplicates and deleted files lingering. - Newly created paths can be found after refresh
exists(path, refresh=True) and invalidatecache(path) now clear cached “not found” results.
Why it matters
The change to job metrics is a breaking default for anyone who expected a stream. Developers should check scripts and integrations that call hf jobs stats or fetchjobmetrics; add the follow option where they need live monitoring. The notes also flag a change to loadtorchmodel: a missing checkpoint path now raises FileNotFoundError rather than ValueError.
The download and cache fixes matter in less dramatic, more useful ways: fewer inconsistent downloads and fewer confusing answers from a filesystem cache that has not caught up with recent changes.
Our read
This is maintenance with direct consequences, not a shiny new model announcement. The snapshot default is sensible for one-off checks and automation, but it changes existing behaviour, so a quick review of monitoring code is worthwhile. The fixes to downloads and cache refreshes are the sort of improvements most people notice only when they are missing. A small release, then, with one change you should not let pass unnoticed.
What to watch
- Whether existing scripts need -f or follow=True to preserve live monitoring.
- Whether downstream tools handle the changed exception type for missing checkpoints.
- Whether the download and cache fixes resolve the reported edge cases in everyday use.
Discussion spark: Should command-line tools default to a quick snapshot for predictable scripts, or preserve live-stream behaviour to avoid quietly breaking existing workflows?
Sources and evidence
- [v2.2.0] hf jobs stats snapshots, download and HfFileSystem fixes (8 October 2026, 15:12 UTC)
not affiliated with or endorsed by Hugging Face