We test whether two sequences are generated by the same distribution or by two different ones. Unlike previous work, we make no assumptions on the distributions' support size. Additionally, we compare our performance to that of the best possible test. We describe an efficiently-computable algorithm based on pattern maximum likelihood that is near optimal whenever the best possible error probability is < exp(-14n2/3) using length-n sequences.