I've run their distilled 70B model and didn't come away too impressed -- feels similar to the existing base model it was trained on, which also rivaled GPT4