This organization hosts pre-converted Keras 3 weights for kerasformers. Each repo holds the converted model plus its tokenizer and config files, so you download once and load directly, no on-the-fly conversion, no transformers or torch runtime on the model path.

Every model is a pure-Keras port that runs unchanged on all three backends. Set KERAS_BACKEND before importing Keras and the same code runs on JAX, PyTorch, or TensorFlow.

Quick start

pip install kerasformers

import os
os.environ["KERAS_BACKEND"] = "torch"   # or "jax" / "tensorflow"

from kerasformers.models.qwen3 import Qwen3Generate, Qwen3Tokenizer

model = Qwen3Generate.from_weights("qwen3-8b")
tokenizer = Qwen3Tokenizer.from_weights("qwen3-8b")

inputs = tokenizer("The capital of France is")
output = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(output[0]))

What's here

Language, vision, audio, and multimodal models, all pure Keras 3:

Browse the models on this page, or see the full catalog in the documentation.

Links

• Library: github.com/IMvision12/KerasFormers
• Documentation: imvision12.github.io/KerasFormers