KerasFormers

KerasFormers is an open-source library that reimplements popular model architectures — from object detectors like DETR to LLMs like Qwen3 and speech recognition with Whisper — entirely in Keras 3. The project has ported over 118 model families, each with weights converted from the original checkpoints. The pitch is simple: you write your code once, and it runs on JAX, PyTorch, or TensorFlow without needing Hugging Face's transformers or torch at inference time.

Getting started is refreshingly straightforward. Install the package with pip, then call `from_weights` on any model class — pass a repo identifier from the KerasFormers Hugging Face organization, a bare variant name, or even a compatible Hugging Face repo with an `hf:` prefix. The corresponding processor is built the same way, so resolution and normalization always match the checkpoint. That's it: two function calls to get predictions. The library includes models for object detection, semantic segmentation, depth estimation, promptable segmentation (SAM 3), open-vocabulary detection (OWLv2), and text generation.

KerasFormers is built for developers and researchers who want to experiment with multiple backends without rewriting pipelines. If you need to compare inference speed across JAX, PyTorch, and TensorFlow, or if you prefer Keras’s API but need access to modern architectures, this library saves you the porting work. It's also handy for production deployments where you want to avoid pulling in the full transformers stack. The project is hosted on GitHub and the weights live on Hugging Face, so integration into existing workflows is about as painless as it gets.