huggingface_hub is not legacy

huggingface_hub is not legacy | Generated by AI

Home 2026.06

Question: Is huggingface_hub legacy now? Is hf the latest CLI? (I have hf --version → 1.12.0)

Answer:

Not legacy — you’re conflating two things. huggingface_hub is the Python package, and it’s very much current. hf is just the new name for its CLI entry point, which ships inside that same package.

What actually changed:

The old command was huggingface-cli. It was renamed to hf (cleaner verb-noun syntax) and the old name was deprecated.
hf --version 1.12.0 is the version of the huggingface_hub package itself — they share versioning because the CLI is part of the library.

So:

pip show huggingface_hub   # this is what hf --version reports
hf --help                  # new CLI (huggingface-cli still works but warns)

The new syntax is hf <resource> <action>:

hf auth login              # was: huggingface-cli login
hf download <repo>         # was: huggingface-cli download <repo>
hf upload <repo> <path>    # was: huggingface-cli upload
hf auth whoami
hf repo create <name>

In Python nothing changed — you still import the same thing:

from huggingface_hub import HfApi, snapshot_download, hf_hub_download

Worth verifying the exact rename timeline and which version flipped the default, since CLIs like this move fast and I shouldn’t guess from memory:Confirmed by the sources — and your 1.12.0 tells the whole story:

The hf CLI was renamed from huggingface-cli as a quality-of-life improvement, introduced in v0.34, following the hf <resource> <action> grammar.
In huggingface_hub v1.0, the deprecated huggingface-cli was fully removed and the redesigned hf CLI (now Typer-based) replaced it. Other v1.0 breaking changes worth knowing if you’re scripting against it: migration to httpx as the HTTP backend, and hf_xet became the default file-transfer package, replacing hf_transfer which was removed.
One compatibility gotcha for your training stack: transformers v4 explicitly requires huggingface_hub v0.x, while transformers v5 requires v1.x. Since you’re on 1.12.0, you need transformers v5 — if something pins old transformers, you’ll get a resolver conflict.

So: huggingface_hub = the library (current, you’re on 1.12.0). hf = its CLI (current). huggingface-cli = the dead name. You’re fully up to date.

Quick reference for the commands you’ll actually hit on the sec-edgar GPT work:

hf auth login
hf download <org>/<model>           # pulls weights/tokenizer
hf upload <org>/<repo> ./out        # single-commit
hf upload-large-folder <repo> ./ckpt  # resumable, for big checkpoints
hf cache ls --filter "size>1GB"     # new v1.0 docker-style cache mgmt
hf jobs run --flavor a10g-small ...  # run scripts on HF hardware

One env var worth setting on your MI300X box to kill the startup auto-update check in non-interactive runs: export HF_HUB_DISABLE_UPDATE_CHECK=1.

References:

Back Donate