Tabu Search (TS) algorithm is one of optimization search methods and has been applied for a large number of combinatorial optimization problems. In the paper, an improved TS algorithm is proposed for protein 3D folding structure prediction in off-lattice AB model. Experimental results show that the lowest energies computed by the improved TS algorithm are better than those obtained by the previous methods. Given a protein sequence, its lowest-energy conformation obtained by our improved TS method forms a single hydrophobic core, which suggests that AB model in three dimensions appears to reflect the real protein reasonably. Compared with the previous heuristic approaches, the improved TS algorithm has higher performance and can be effectively used to predict 3D structure prediction of proteins.