I've been using a finetuned T5 model for commercial text generation for the past six months. The volume of discussion around the model on github and elsewhere leads me to believe others are as well, although people tend to be circumspect about implementation details.
How are you using it, through huggingface or directly?
Check out the T5 models at huggingface for this. Main NLG use cases with T5 are translations, summarization and question generation. Latter is sophisticated + nothing trivial and "NLG", so yeah.
From my experience T5 is currently the best publicly available model for NLG.
Did you use it already in some production environment?
Not yet but I am working on it!
Nice, will you use it via huggingface?
Yeah currently experimenting with T5 large in FP16. Will then use the ONNX conversion script to decrease inference time and deploy it!