Transformers v5 release: PyTorch-first backend, new inference APIs, and transformers serve

AI Impact Summary

Transformers v5 centralizes PyTorch as the sole backend, expands model coverage, and positions transformers serve as an OpenAI API–compatible inference server for large-scale evaluation. The release introduces new inference abstractions (continuous batching, paged attention) and standardizes tokenization backends, while moving away from Flax/TensorFlow. This creates a clearer migration path for teams investing in PyTorch-first workflows but requires planning for TF/JAX deprecation and for adopting the new serving model in production. The ecosystem partners cited (Unsloth, Axolotl, LlamaFactory, TRL, MaxText) indicate broad integration potential but also a need to align tooling and CI across multiple frameworks.