Yeah currently experimenting with T5 large in FP16. Will then use the ONNX conversion script to decrease inference time and deploy it!