Factoring out nucleotide-level mutation biases from antibody language models dramatically improves prediction of functional mutation effects while reducing computational cost by orders of magnitude.