When applying deep neural networks (DNNs) in resource-constrained environments, Lightweight model techniques are often used to modify the model to be less computationally intensive. However, some of these techniques can struggle to balance maintaining their high accuracy against reducing processing costs. Lightweight branching models add early exits to a model for quicker inference, but these earlier exits can suffer a reduced level of accuracy because of the lack of learned knowledge in the preceding layers. We propose the use of a novel form of Knowledge Distillation to assist in the branched model’s training by distilling knowledge from the main classifier exit to the branch exits. This novel technique, entitled branching self-distillation, combines the student-teacher training of knowledge distillation with the effective efficiency optimisation of branching to create a novel lightweight optimisation technique that requires minimal additional training to achieve improvements in both accuracy and processing costs. We demonstrate the effectiveness of the technique and achieve an average of 1.86% increase in accuracy and a 38.02% further reduction in average processing flops per input upon a selection of well-known lightweight model architectures. (Source code and further experiment details can be found online at this work’s repository at https://github.com/SanityLacking/Self-Distillation ).