Recent efforts for democratising protein structure prediction have leveraged the MMseqs2 algorithm to efficiently generate multiple sequence alignments with high diversity and a limited number of sequences. Here, we investigated the usefulness of this strategy for mutational outcome prediction. We place ourselves in a context where we only exploit information coming from the input alignment for making predictions. Through a large-scale assessment of ~1.5M missense variants across 72 protein families, we show that the MMseqs2-based protocol implemented in ColabFold compares favourably with tools and resources relying on profile-Hidden Markov Models. Our study demonstrates the feasibility of simultaneously providing high-quality and compute-efficient alignment-based predictions for the mutational landscape of entire proteomes.