Hugging Face

Models and offerings #

Who they are

This hub is for anyone who downloads a model, which makes it the most widely used platform on this site. Hugging Face started in 2016 as a chatbot company, published pytorch-pretrained-bert in late 2018 when BERT appeared, and turned that into Transformers. The Hub followed in 2020 as a Git-based host for weights, and the company has since positioned itself as the neutral meeting point between model makers (Meta, Google, Mistral, Alibaba, NVIDIA and thousands of individuals) and the people who run their models.

Its place in the field is the distribution layer. A model is “released” when it is on the Hub, a dataset is “public” when it is on the Hub, and fine-tuning recipes are written against Transformers, PEFT and TRL. Hugging Face does not train frontier models itself, apart from small open ones such as the SmolLM series, so it has no competing interest in which model you pick.

Tooling and APIs

pip install transformers datasets accelerate peft trl gets you the stack. Transformers v5.0.0 (26 January 2026) was the first major release in five years: it dropped the TensorFlow and Flax model classes and is PyTorch only, from_pretrained now loads in the dtype the weights were saved in rather than float32, each model has one tokeniser file backed by the fastest available backend, and apply_chat_template returns a BatchEncoding. Releases come roughly weekly, with v5.23.0 on 13 November 2026 adding six new architectures. The minimum Python is 3.10. Everything runs on CPU, CUDA, ROCm and Apple MPS through PyTorch, and the Apple hub covers the MLX conversions that live under the mlx-community organisation.

Loading a small instruction model takes four lines:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

torch.manual_seed(0)
name = "HuggingFaceTB/SmolLM2-135M-Instruct"
tokenizer = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name)

messages = [{"role": "user", "content": "Name one use of a linear regression model."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)
out = model.generate(**inputs, max_new_tokens=24, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

On our machine this prints One use of a linear regression model is to predict the value of a continuous outcome variable based on one or more predictor variables, in under three seconds on CPU once the 135 M parameter weights are cached. SmolLM2-135M-Instruct is 135 M parameters, 8 k context, Apache-2.0, released November 2024.

Pricing, at the time of writing and in US dollars: the Hub, the libraries, public and private repositories within quota, Spaces on a basic CPU and ZeroGPU time are free. PRO is 9 USD a month and includes 2 USD of Inference Providers credit, Team is 20 USD per user a month, Enterprise 50 USD per user a month with SSO, SCIM and audit logs. Inference Endpoints are billed by the hour: a CPU instance from 0.03 USD, an L4 at 0.80 USD, an A100 80 GB at 2.50 USD, an H100 at 4.50 USD, a B200 at 9.25 USD. Storage beyond the free quota is 12 USD per TB per month for public repositories and 18 USD for private. Prices and quotas change, so check huggingface.co/pricing.

Pain points: Transformers is large (hundreds of architectures) and import time shows it, generation defaults differ per model and are buried in generation_config.json, and the Hub cache under ~/.cache/huggingface grows without bound. Text Generation Inference, the company’s own serving stack, is in maintenance mode according to its repository, which points users to vLLM, SGLang and llama.cpp instead.

Migration notes

From raw PyTorch: a Transformers model is an nn.Module with a config object, so you can drop into the forward pass or fine-tune with your own loop. The Trainer class and accelerate launch cover the common distributed cases. The LoRA fine-tune project uses PEFT this way.

Model formats: safetensors is the Hub default and loads anywhere. To serve a Hub model you usually do not convert it: vLLM and TensorRT-LLM read Hub repositories directly. For llama.cpp or Ollama you convert to GGUF with the scripts in the llama.cpp repository, and many organisations publish GGUF and MLX variants alongside the originals. The quantisation formats reference maps these names.

From a hosted API: InferenceClient and the Inference Providers router speak an OpenAI-compatible chat completions format, so code written for OpenAI usually needs only a base URL and a model id change.

Watch for

The library is Apache-2.0 but the models are not. Every repository has its own licence field and many are gated: Llama and earlier Gemma checkpoints require accepting the maker’s terms before the download works, and some research models are non-commercial. Read the licence before you fine-tune, because derivatives inherit it. The models reference records the licence of each model we track.

trust_remote_code=True executes Python from the repository you are loading. Use it only for organisations you trust and pin a revision. Tokens created at huggingface.co are bearer credentials: keep them in HF_TOKEN, never in a notebook. The Hub and the libraries collect usage telemetry by default (download counts and library versions). Set HF_HUB_DISABLE_TELEMETRY=1 and HF_HUB_OFFLINE=1 in locked-down environments, which is also how the Field Guide pipeline runs.

Rate limits apply to anonymous downloads and to the free Inference Providers tier, and the free credit is small. Inference Endpoints bill while an endpoint is running even if idle, so set scale-to-zero. Transformers v5 broke pickled tokenisers and some custom code written for v4, and a weekly release cadence means examples on the Hub drift out of date within months.