Overview
Transformers 5.0, released on December 18, 2025, goes PyTorch-only, adopts a modular architecture and redesigns tokenizers.
Main Features
PyTorch-only
Transformers 5.0 drops TensorFlow and JAX backends to focus on PyTorch, simplifying maintenance and testing.
python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('gpt2')
tokenizer = AutoTokenizer.from_pretrained('gpt2')
inputs = tokenizer('Hello', return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0]))Modular architecture
Models are now organized into reusable modular components, making it easier to create new architectures by composition.
Tokenizer redesign
Tokenizers have been redesigned with a unified API and improved performance thanks to tighter integration with the Rust-based tokenizers library.