KerasFormers

I started KerasFormers to solve a problem I kept running into: many modern models were available in Hugging Face, but using them in Keras 3, especially across TensorFlow, JAX, and PyTorch backends, often meant writing or adapting a lot of model-specific code.

What started as a small experiment grew into a library covering 60+ vision, speech, LLM, and VLM architectures, with Hugging Face to Keras weight conversion and parity validation built in.

The biggest challenge was making the implementations genuinely backend-agnostic rather than just wrapping existing PyTorch code. That meant digging into model architectures, tensor layouts, Keras 3 ops, checkpoint formats, and numerical parity.

KerasFormers is still evolving, and I’m continuing to add models, improve conversions, and make the ecosystem easier to use. I’d love feedback on what models or capabilities would be most useful next.