Adept's 8B parameter model with novel architecture improvements for efficient training and inference.