Transformers

Getting Started

1. Install the library — Run pip install transformers to install the package, optionally alongside torch, tensorflow, or jax depending on your preferred backend.
2. Pick a model — Browse the Hugging Face Hub to find a pretrained model suited to your task (e.g., meta-llama/Llama-3, openai/whisper-large, google/vit-base-patch16-224).
3. Run inference — Use the pipeline() API for quick inference: from transformers import pipeline; pipe = pipeline('text-generation', model='gpt2'); pipe('Hello, world!').
4. Fine-tune or customize — Load model weights with AutoModel and AutoTokenizer classes and plug them into your own training loop or use the built-in Trainer API.

Key Features

• Unified Model Hub — Access over 500,000 pretrained model checkpoints directly from the Hugging Face Hub with a single line of code.
• Latest release (early March 2026) — The most recent release adds support for new model architectures, hardware backends, and continued refinements to the pipeline API.
• Multi-framework support — Works seamlessly with PyTorch, TensorFlow, and JAX, with easy model conversion between backends.
• Pipeline API — High-level pipeline() abstraction lets you run inference for text generation, classification, summarization, translation, image recognition, and more without boilerplate.
• Fine-tuning tools — Includes a Trainer class and integration with PEFT, making it straightforward to fine-tune large models with minimal memory overhead.
• Broad modality coverage — Supports text, vision, audio, video, and multimodal models under a single consistent API.
• Latest release (early March 2026) — The most recent release continues to expand model coverage, framework compatibility, and inference performance.
• Local and private inference — Models run entirely on your own hardware with no data sent to external servers, suitable for sensitive or offline workloads.
• Latest release (early March 2026) — The most recent release continues to expand model compatibility, hardware optimizations, and support for the latest architectures on the Hub.