transformers 5.15.1

The most important sentence in the transformers README is buried in the 'when not to use' section. It's still there in 5.15.1: the model files are deliberately not refactored into modular building blocks, so researchers can iterate on a single architecture without navigating layers of abstraction. That design choice is the whole philosophy. Every other decision flows from it.

What transformers actually does is serve as the canonical model-definition layer for the broader ML ecosystem. The README is explicit about this: when a model is defined here, it becomes compatible with training frameworks like Axolotl, Unsloth, DeepSpeed, and FSDP, and with inference engines like vLLM, SGLang, and TGI. The library is the agreed-upon specification, not just a convenient wrapper. That's a different value proposition than 'easy pretrained models,' and it's worth taking seriously.

Over 1M+ checkpoints on the Hub now carry the transformers tag. The modality coverage has grown well past NLP: the example list spans audio (Moshi for speech-to-speech, CSM for text-to-speech, Moonshine for streaming ASR), computer vision (DepthPro, VitPose, SuperGlue for keypoint matching), and multimodal work including Voxtral for audio-or-text-to-text and GOT-OCR2 for document understanding. The breadth is real, not aspirational—each entry links to an actual Hub model page.

The minimum requirements are stated plainly: Python 3.10+ and PyTorch 2.5+ are now the floor. That's a meaningful constraint for anyone running older infrastructure.

The Pipeline API remains the entry point for inference, and the README now surfaces a CLI path too—transformers chat against any instruction-tuned model, provided transformers serve is running. That's a small but telling addition: the library is increasingly positioning itself as something you run, not just something you import.

The honest limitation the README names is that the training API is optimized specifically for PyTorch models provided by transformers itself. Generic training loops belong elsewhere, in Accelerate or similar. The example scripts are explicitly labeled as starting points that will need adaptation.

For agent builders specifically, the ecosystem-pivot framing matters. If you're choosing which model definition to trust as stable across your inference engine, your fine-tuning framework, and your quantization toolchain, the answer this library is giving is: use ours, because everyone else already does.