GoMLX

Supports OpenXLA for CPU/GPU/TPU acceleration with state-of-the-art speed comparable to Jax and TensorFlow, as noted in the README, while also offering a pure Go backend for portability.

Runs almost everywhere Go runs, including browsers via WASM, enabling ML in web applications without C++ dependencies, as demonstrated in the Hive game demo.

Can convert and run ONNX models from HuggingFace using the onnx-gomlx package, facilitating easy use of pre-trained models like Gemma 3 and BERT for inference or fine-tuning.

Includes autodiff, common layers (FFN, convolutions, transformers, KANs), training tools, optimizers, and loss functions, providing a full-featured platform for building and experimenting with models.