The bitnet paper was showing worse results than fp16 transformer with the same parameter count. The shocking result in the 1.58b paper (same group) is no quality loss compared to fp16.