Bioluminescence tomography (BLT) is a new optical molecular imaging technique that utilizes the light intensity information on the surface of biological tissues to reconstruct the three-dimensional distribution of the internal bioluminescent source. The source reconstruction of BLT has serious morbidity due to the limited measurements of light intensity and the complicated structure of biological tissue. A multilevel adaptive finite element method for BLT is proposed, which is combined with the permissible regional-shrinking strategy to improve the quality of the reconstruction. Simulations of single-source and double-source based on the digital mouse model are designed to assess the localization ability of light source and the quantification ability of energy density of the method, respectively. The results show that the proposed method can significantly improve the positioning accuracy and energy density of the light source.