Hugging Face’s Transformers library is the default model toolkit for much of the AI world. Its first release candidate of version 5 arrived in December 2025, five years after the v4 candidate, and the stable release followed in early 2026. The announcement noted that the library was being installed more than 3 million times a day, up from 20,000 daily installs in the v4 era, with over 1.2 billion installs in total.
What is new
The theme is interoperability. Transformers v5 aims to serve as the shared source of truth for model definitions that other engines such as vLLM, SGLang and llama.cpp can build upon. A shared attention interface replaces much per-model custom logic, and the library gains inference features including continuous batching, paged attention and a transformers serve command that exposes an OpenAI-compatible API.
Reports also cite large speedups for Mixture-of-Experts models, and a simplified tokenizer system that removes the old split between slow and fast tokenizers.
What was removed
PyTorch is now the primary backend, and TensorFlow and Flax support was dropped. Teams still relying on those frameworks need a migration plan. Hugging Face also moved to weekly minor releases so that new models land faster.
A practical warning
The tokenizer rework has caused problems for some older checkpoints, particularly SentencePiece-based ones, and unpinned notebooks that install the latest version have broken as a result. The lesson is old but important: pin your dependency versions, read the migration guide and test before you upgrade.