Sori-1B is a 1B-parameter audio-language model built by a single SNU researcher whose core claim to fame is that its decoder is trained entirely from scratch on audio-paired text — no text-only pretraining, no pretrained-LM initialization — with the idea being that a language model which has only ev
