Modern phylogenomic studies now commonly analyse hundreds or thousands of loci, yet traditional tree inference methods assume that each locus or partition evolves under a single model of sequence evolution and shares the same tree topology. However, this is far removed from reality, especially as different loci often exhibit distinct evolutionary histories due to processes such as incomplete lineage sorting (ILS), introgression, or horizontal gene transfer. To address these limitations, mixture models have been introduced in IQ-TREE2, an efficient software for phylogenomic inference based on maximum likelihood. We present three contributions: (1) an algorithm called MixtureFinder, which selects the optimal number of Q-matrix classes for modelling DNA sequence evolution; (2) a mixture across sites and trees (MAST) model that allows more than one tree along an alignment; and (3) a Hidden Markov Model (HMM) that consider intercorrelations between neighboring sites of an alignment. Simulations and empirical data analyses demonstrate that these approaches can capture and identify biological processes, suggesting that they will be useful for future empirical studies.