Bioluminescence tomography (BLT) is a promising imaging modality that can provide noninvasive three-dimensional visualization information on tumor distribution. In BLT reconstruction, the widely used methods based on regularization or greedy strategy face problems such as over-sparsity, over-smoothing, spatial discontinuity, poor robustness, and poor multi-target resolution. To deal with these problems, combining the advantages of the greedy strategies as well as regularization methods, we propose a hybrid reconstruction framework for model-based multispectral BLT using the support set of a greedy strategy as a feasible region and the Alpha-divergence to combine the weighted solutions obtained by $L$1-norm and $L$2-norm regularization methods. In numerical simulations with digital mouse and in vivo experiments, the results show that the proposed framework has better localization accuracy, spatial resolution, and multi-target resolution.Bioluminescence tomography (BLT) is a promising imaging modality that can provide noninvasive three-dimensional visualization information on tumor distribution. In BLT reconstruction, the widely used methods based on regularization or greedy strategy face problems such as over-sparsity, over-smoothing, spatial discontinuity, poor robustness, and poor multi-target resolution. To deal with these problems, combining the advantages of the greedy strategies as well as regularization methods, we propose a hybrid reconstruction framework for model-based multispectral BLT using the support set of a greedy strategy as a feasible region and the Alpha-divergence to combine the weighted solutions obtained by $L$1-norm and $L$2-norm regularization methods. In numerical simulations with digital mouse and in vivo experiments, the results show that the proposed framework has better localization accuracy, spatial resolution, and multi-target resolution.