dev-newsHugging Face logohuggingface.co

Latest stable huggingface_hub: 2.3.0, released (release notes). 137 stable releases since 5 Jan 2024.

huggingface_hub

38minor releases in the 53 weeks since 1.0

1.0.0.01.1.0.1Nov1.2.0.2Dec1.3.0.3Jan1.4.0.4Feb1.5.0.51.6.0.6Mar1.7.0.71.8.0.81.9.0.91.10.0.10Apr1.11.0.111.12.0.121.13.0.131.14.0.14May1.15.0.151.16.0.161.17.0.171.18.0.18Jun1.19.0.191.20.0.201.21.0.211.22.0.221.23.0.23Jul1.24.0.241.25.0.251.26.0.261.27.0.27Aug1.28.0.281.29.0.291.30.0.301.31.0.31Sept1.32.0.321.33.0.332.0.02.02.1.02.12.2.02.2Oct2.3.02.317 weeks in a row1.x on httpx2.x on httpx2
One column per week from 20 October 2025, from the GitHub releases of huggingface/huggingface_hub. Rungs are minor releases (.8 means 1.8, white rungs are 1.0 and 2.0); dots underneath are patch releases.

huggingface_hub 2.0.0 came out on 24 September 2026, eleven months after 1.0.0 dropped requests for httpx. 2.0 swaps the HTTP stack again, from httpx to httpx2, so custom client factories and except clauses for transport errors have to use httpx2 types; code that must run on both lines can import httpx from huggingface_hub.utils, available since 1.30.0. The release also removes every API deprecated in 1.x, including upload_large_folder and the huggingface-cli entry point, now hf.

The 1.x line ended at 1.33.0, published under two hours before 2.0.0 on the same day. From 23 March a new minor shipped every week for 17 weeks, 1.8 through 1.24. 2.x has kept that pace with 2.1 on 1 October, 2.2 on 8 October and 2.3 on 22 October, which lets HfFileSystem follow Bucket changes live instead of serving stale directory listings.

As of 24 October 2026. huggingface_hub 2.0.0 release notes.

Latest releases

  1. 2.3.0HfFileSystem can follow Bucket changes live with live_follow=True.
  2. 2.2.0hf jobs stats prints one snapshot and exits; downloads reliably use hf_xet.
  3. 2.1.1Renames hf spaces zero-gpu-quota to hf spaces zero-gpu quota.
  4. 2.1.0Jobs gain retries, reruns, rescheduling and exposed ports.
  5. 2.0.0HTTP stack moves to httpx2; APIs deprecated in 1.x removed.
  6. 1.33.0Validating progress bar for large Xet uploads.

Moving to huggingface_hub 2.0

2.0.0 (24 September 2026) has two breaking changes: the HTTP stack moves from httpx to httpx2, and every API deprecated during 1.x is gone. Code that only calls hf_hub_download, snapshot_download or HfApi methods that still exist needs no change.

  1. If you catch transport errors or pass your own client factory, import the HTTP module from huggingface_hub.utils. That works on 1.30.0 and later as well as 2.x:
    from huggingface_hub.utils import get_session, httpx

try:
    r = get_session().get("https://huggingface.co/api/models/gpt2")
    r.raise_for_status()
except httpx.HTTPError:
    ...
  2. Replace removed calls: upload_large_folder becomes upload_folder, duplicate_space becomes duplicate_repo, list_models(model_name=...) becomes list_models(search=...), and text_generation(stop_sequences=...) becomes stop.
  3. Replace removed commands in scripts and CI: huggingface-cli becomes hf, hf repo becomes hf repos, and hf upload-large-folder becomes hf upload.

httpx2 uses the operating system's certificate store by default; SSL_CERT_FILE and SSL_CERT_DIR still work. Logging configuration has to target the httpx2 and httpcore2 loggers.

2.0.0 release notes Migration guide

Access tokens for git and the hf CLI

The Hub takes a User Access Token wherever it used to take a password: for git over HTTPS, for basic auth, and in the Python libraries. Tokens are read, write or fine-grained; Hugging Face recommends fine-grained tokens for production and one token per machine or app, so one can be revoked without breaking the others.

  1. Create a token on the Access Tokens tab of your Hub settings, then log in. hf auth login is interactive; for scripts pass the token from an environment variable, and add it to git's credential helper if you push with git:
    hf auth login
hf auth login --token $HF_TOKEN --add-to-git-credential
hf auth whoami
  2. In CI, Trusted Publishers swaps the CI provider's OpenID Connect token for a short-lived Hub token on each run, so no long-lived token has to be stored as a secret.

If a token leaks, refresh or delete it on the Access Tokens tab. A token you find that belongs to someone else can be invalidated by anyone with POST https://huggingface.co/api/credentials/revoke; the owner gets an email. The endpoint answers 202 whether or not the token existed.

User access tokens hf CLI guide

RLOO training with TRL

RLOO (REINFORCE Leave-One-Out) is an online RL method in TRL, Hugging Face's post-training library. For each prompt the model samples several completions; each one's advantage is its reward minus the mean reward of the other completions for the same prompt. That baseline stands in for the value model PPO trains, so there is one model fewer in memory. The method is from Ahmadian et al., "Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs" (2024).

rewardadvantage0.90.37+0.53o10.20.60−0.40o20.60.47+0.13o30.30.57−0.27o4
Four completions of one prompt, example rewards. The white tick on each bar is its baseline, the mean reward of the other three: for o1, (0.2 + 0.6 + 0.3) ÷ 3 = 0.37. The advantage below is reward minus baseline, so the completions that beat their siblings are pushed up and the rest down.
  1. A minimal run with the current RLOOTrainer takes a model id, one or more reward functions and a dataset with a prompt column. Start it with accelerate launch train_rloo.py:
    from datasets import load_dataset
from trl import RLOOTrainer
from trl.rewards import accuracy_reward

dataset = load_dataset("trl-lib/DeepMath-103K", split="train")
trainer = RLOOTrainer(
    model="Qwen/Qwen2-0.5B-Instruct",
    reward_funcs=accuracy_reward,
    train_dataset=dataset,
)
trainer.train()
  2. Reward functions are plain Python callables (sync or async def) that take prompts, completions and **kwargs and return one float per completion, so you don't need a trained reward model.
  3. Set use_vllm=True in RLOOConfig to generate with vLLM, either colocated on the training GPUs (the default) or as a separate server with vllm_mode="server".

TRL 0.22.0 (29 August 2025) rewrote RLOOTrainer to match GRPOTrainer. Configs from 0.21 and earlier need renames: rloo_k is now num_generations, kl_coef is beta, cliprange is epsilon, num_ppo_epochs is num_iterations, and the trainer's config, policy and reward_model arguments became args, model and reward_funcs. The latest TRL on PyPI is 1.16.0, from 15 October 2026.

TRL RLOO trainer docs RLOO paper TRL on PyPI

Questions people search for

What is the latest version of huggingface_hub?

huggingface_hub 2.3.0, released 22 October 2026 (release notes). Install or upgrade it with pip install -U huggingface_hub.

Is huggingface-cli still available?

Not in huggingface_hub 2.0, which removed the huggingface-cli entry point; use hf instead, for example hf auth login or hf download. It was deprecated during 1.x.

How do I log in to Hugging Face from the command line?

Run hf auth login and paste a User Access Token from your Hub settings. In scripts, use hf auth login --token $HF_TOKEN, adding --add-to-git-credential if you push with git. hf auth whoami shows which account you are logged in as.

What is RLOO in TRL?

REINFORCE Leave-One-Out: for each prompt the model samples several completions, and each completion's advantage is its reward minus the average reward of the others. TRL implements it as RLOOTrainer; it needs no value model, unlike PPO.