Hacker News Reader

In case anyone is wondering, this repo uses various fine-tuned Flan-T5 models.

And Flan-T5 models are awesome! Fastchat is a finetuned 3b parameter Flan-T5 model that can perform really well, comparably to llama 13b

That's correct. The current base model is an int8 quantization of LaMini-Flan-T5-248M described here:

I shared more details over on Reddit: