Transformers

Hugging Face's Transformers library holds the reference Python implementation for a huge range of model architectures, from BERT and T5 through Llama, Qwen and Gemma to Whisper, SAM and vision-language models. Each architecture lives in a mostly self-contained file, and weights load from the Hub or a local folder with one from_pretrained call. The pipeline API covers quick inference for text generation, speech recognition, image classification and more, the Trainer class handles fine-tuning, and the transformers command adds a terminal chat client and a serve mode that answers OpenAI-style requests. Because Axolotl, Unsloth, vLLM, SGLang and other tools reuse these definitions, a model added here usually works across that ecosystem. Version 5 dropped TensorFlow and JAX to focus on PyTorch 2.5 or newer, with Python 3.10 or later; install with pip and the torch extra. Small models run on CPU, larger ones need a GPU. The code is Apache 2.0, but each checkpoint has its own license, so read the model card before shipping. Suited to ML engineers and researchers working across modalities.