A new model reduction method is presented for discrete delay systems, where the reduced system is also a discrete delay system of a smaller size in the dimension of the state. Traditional model reduction approaches to discrete delay systems are based on transforming delay systems into non-delay equivalent systems. However, when the state delay is large, they are computationally demanding. The proposed model reduction method can be efficiently computed irrespective of the state delay. To do this, a diagonal balanced realization is proposed based on block diagonal gramian-like matrices, and properties of truncation of the diagonal balanced realization are investigated.