Hugging Face logo, a yellow hugging face emoji next to the Hugging Face wordmark, for the hf CLI new features guide

hf CLI New Features: Power-User Guide to huggingface_hub 2.3 and 2.4, Jobs, Sandboxes and Buckets

by

Last updated: October 30, 2026

The hf CLI new features from October 2026 include a shared cache that downloads each file once across repos, Jobs retries and reruns, faster Storage Bucket downloads, experimental cloud Sandboxes, and an agent skill. huggingface_hub 2.3.0 and 2.4.0 add live Bucket listings, parallel repo listings, Bucket file copies and extension update checks.

This guide draws on Hugging Face’s October 26, 2026 post What’s new in the hf CLI and huggingface_hub library, the v2.3.0 and v2.4.0 release notes, and the official CLI guide. PyPI lists 2.3.0 on October 23 and 2.4.0 on October 29, 2026.

Contents

hf CLI new features in October 2026

The hf command and the huggingface_hub Python library ship together, so a library release also changes the CLI. Hugging Face says it has been shipping features “these past few weeks” across Jobs, Storage Buckets, Sandboxes, cache optimization and downloads and uploads.

Agents are now a major audience. According to the recap, on a given day one in three users of the hf CLI is an AI agent.

Feature summary

Feature What it does Release
Shared Xet cache Each Xet file is stored once; other repos get a link instead of a new download Recent releases (Oct 26 recap)
Cache cleanup hf cache rm for a single file, hf cache prune for unused shared files Recent releases
Jobs --attempts, hf jobs rerun, hf jobs scheduled reschedule, hf jobs expose, automatic names Recent releases
Webhook-triggered Jobs A repo or Bucket change starts a Job, with the event in WEBHOOK_PAYLOAD Recent releases
Bucket live listings HfFileSystem(live_follow=True) 2.3.0
Parallel listings find and glob split by sub-directory; max_workers defaults to 16 2.3.0
Attachments HfApi.upload_attachment and --attach on discussions and PRs 2.3.0
Bucket copies cp_file, copy and mv into Buckets 2.4.0
Extension checks VERSION and LATEST columns in hf extensions list 2.4.0
Sandboxes (experimental) hf sandbox create, exec and kill Recent releases

Shared cache and downloads

The same file often appears in many repos. Hugging Face’s example is a 9.5 GB text encoder shipped by FLUX.1-dev, FLUX.1-schnell and FLUX.1-Krea-dev. The cache now stores each Xet file once, and other repos get a link, so no bytes are transferred and no extra disk is used.

Cleanup is also more precise. You can remove one cached file, such as a single GGUF quantization, without deleting the whole repo:

hf cache rm hf://models/unsloth/gemma-3-27b-it-GGUF/gemma-3-27b-it-Q4_K_M.gguf
hf cache prune

Downloads changed too. hf download no longer makes one API call per file, which used to hit rate limits on repos with tens of thousands of files. Parallel downloads always use hf_xet, and the CLI warns you when --include or --exclude patterns match no files. If you pull GGUF files for local runners, this pairs well with our Ollama new features guide and LM Studio Bionic guide.

Jobs: retries, reruns and webhooks

Hugging Face Jobs run a command in a Docker image on Hugging Face hardware. The Jobs guide says they are available to any user or organization with a positive credit balance and are pay-as-you-go by the second. The new commands are:

hf jobs run --attempts 3 python:3.12 python train.py   # retry on failure
hf jobs rerun <job_id>                                  # same secrets and hardware
hf jobs scheduled reschedule <id> "0 9 * * *"           # change the schedule
hf jobs expose <job_id> 8000 --public 9000              # open ports on a running job

A webhook can also start a Job instead of calling a URL, for example hf webhooks create --job-id <job_id> --watch bucket:my-org/raw-data --secrets HF_TOKEN.

For agents, hf jobs stats now prints a snapshot and exits, hf jobs logs without -f no longer hangs, and per the 2.3.0 notes hf jobs run reconnects until the Job finishes and exits immediately on Ctrl+C.

Storage Buckets and HfFileSystem

HfFileSystem now downloads Xet files from Buckets with hf_xet, straight to disk. In Hugging Face’s benchmark on HF Jobs, a 2.15 GB parquet file went from 2 minutes to about 7 seconds.

Version 2.3.0 added live_follow=True, which keeps one background connection per listed Bucket so new files appear in listings without clearing the cache. It is off by default and does not apply to model, dataset or Space repos. The same release lists large repos in parallel: on datasets/allenai/objaverse, a full find() over 798,924 files dropped from 776 seconds to 19.7 seconds.

The 2.4.0 release lets you copy files into a Bucket from the same Bucket, another Bucket, or a model, dataset or Space repo. Xet files are copied on the Hub by hash, so nothing is downloaded. Copying from a Bucket to a repo is not supported yet.

Sandboxes and the hf-cli skill

Sandboxes, marked experimental, are isolated cloud machines you start, run commands in, and move files in and out of. They run on Jobs, so any Docker image works on CPU or GPU, and custom labels help you find the sandboxes a run created or track costs.

The hf-cli skill gives coding agents the CLI’s command surface. The CLI guide says hf skills add works with Claude Code, Codex, Cursor, OpenCode, Pi and any agent that loads skills from .agents/skills. In Hugging Face’s test of 28 Hub tasks, DeepSeek-V4.1-Flash and Kimi-K3 solved 99 to 100 percent with or without the skill, but used about a third fewer turns with it. See our guides to Claude Code, Codex CLI 0.166, DeepSeek V4.1 Flash and Kimi K3.1 Pro.

How to upgrade and try the new features

  1. Install or update. Run hf update, or install fresh with curl -LsSf https://hf.co/cli/install.sh | bash, brew install hf, or pip install -U huggingface_hub. To try it without installing, use uvx hf.
  2. Sign in with hf auth login. For scripts, pass --token, since the login is interactive.
  3. Add the agent skill with hf skills add if you use a coding agent. The standalone installer adds it by default; pass --exclude-skill to skip it.
  4. Reclaim disk space with hf cache rm for single files and hf cache prune for unused shared files.
  5. Make long Jobs resilient with --attempts, and use hf jobs rerun to repeat a Job with the same secrets and hardware.
  6. In Python, create HfFileSystem(live_follow=True) when another process writes to a Bucket you are reading.
  7. Run hf extensions list to see whether any CLI extension is outdated, then hf extensions update.

Limitations

  • Sandboxes are experimental.
  • Jobs and Sandboxes need a positive credit balance and are billed per second.
  • live_follow works only for Buckets and keeps a connection open per followed Bucket.
  • Copying from a Bucket to a repo is not supported yet.
  • Extension update checks query GitHub at most once a day; set HF_HUB_DISABLE_UPDATE_CHECK=1 to turn them off.

Frequently asked questions

What are the hf CLI new features in October 2026?

Hugging Face’s October 26, 2026 recap lists a shared Xet cache that stores each file once across repos, Jobs retries, reruns and rescheduling, faster Storage Bucket downloads with live listings, experimental Sandboxes, an hf-cli skill for AI agents, and fewer API calls during hf download.

What is the latest version of huggingface_hub?

PyPI lists huggingface_hub 2.4.0, published October 29, 2026. It follows 2.3.0, published October 23. Version 2.4.0 adds file copies into Storage Buckets and shows when hf CLI extensions are outdated.

How do I update the hf CLI?

Run hf update. It detects whether hf was installed with Homebrew, the standalone installer or pip and runs the matching update command. If the hf-cli skill is installed globally, it is refreshed too. You can also run pip install -U huggingface_hub.

What are Hugging Face Sandboxes?

Sandboxes are an experimental feature: isolated cloud machines you start from Python or the CLI, run commands in, and move files in and out of. They run on Hugging Face Jobs, so any Docker image works on CPU or GPU.

Who can use Hugging Face Jobs?

Hugging Face says Jobs are available to any user or organization with a positive credit balance. Jobs are pay-as-you-go, so you pay only for the seconds you use, and you choose hardware with the –flavor option.

Sources

Image: Hugging Face logo with title, from Wikimedia Commons (public domain; originally from Hugging Face brand assets), placed unchanged on a white background. Hugging Face is a trademark of Hugging Face, Inc. FreeAnswers is not affiliated with or endorsed by Hugging Face.



Related answers

Leave a Comment