Source　code　summaries　provide　concise　natural　language　descriptions　aiding　in　software　maintenance.　In　traditional　deep　learning-based　code　summarization　techniques,　a　single　model　with　the　same　architecture　is　typically　used　to　handle　all　data,　without　considering　the　differences　in　source　code　characteristics.　This　paper　presents　a　model　fusion　strategy　that　integrates　multiple　model　architectures.　This　strategy　dynamically　allocates　weights　through　the　gating　network,　using　different　single　models　for　summary　generation　based　on　the　distinct　characteristics　of　the　code.　It　fully　leverages　the　various　embedding　methods　and　model　structures　of　multiple　single　models,　finding　the　appropriate　processing　combination　for　codes　with　different　characteristics.　The　strategy　harnesses　the　advantages　of　the　supervised　learning　model＇s　explicit　knowledge　learning　and　the　pre-trained　model＇s　deep　semantic　understanding.　Coupled　with　various　code　feature　extraction　methods,　it　effectively　generates　high-quality　summaries　for　code　snippets　of　varying　types　and　complexities.　Two　widely　recognized　public　datasets　are　used　to　train　and　evaluate　our　method.　The　results　show　that　our　model　fusion　strategy　substantially　elevates　code　summarization　quality　over　single　models,　enhancing　semantic　alignment　between　source　code　and　summaries.　©　2023　IEEE.