Transformers v5: the first major release in five years

What happened

Transformers v5.0.0 was released.

The technical details

From the release notes:

• the first major release in five years, with 1,200 commits since the last minor release
• removes many long-due deprecations, and refactors to simplify the APIs and internals
• significant API changes around dynamic weight loading and tokenization
• a continuously updated migration guide
• release cadence changes to weekly minor releases (v5.1, v5.2, …) so new models are available sooner

Why it matters

I felt v5 directly in this blog’s code. apply_chat_template(..., return_tensors="pt") now returns a dict-like object instead of a tensor, so my tool-calling and prompting snippets had to use return_dict=True and ["input_ids"]. Models now print a load report too. Major versions break things, and the migration guide is the first place to look.