Lattice basis reduction is a mandatory tool to solve lattice problems such as the shortest vector problem (SVP), whose hardness assures the security of lattice-based cryptography. The most famous reduction is the celebrated algorithm by Lenstra-Lenstra–Lovász (LLL), and the block Korkine–Zolotarev (BKZ) is its blockwise generalization. At present, BKZ and its variants such as BKZ 2.0 are a de facto standard reduction algorithm to estimate the security level of lattice-based cryptosystems. Recently, DeepBKZ was proposed as a mathematical improvement of BKZ, in which LLL with deep insertions (DeepLLL) is called as a subroutine alternative to LLL. In this paper, we develop a new self-dual variant of DeepBKZ to obtain a reduced basis. Different from conventional self-dual algorithms, we select suitable free dimensions to reduce primal and dual lattice bases in our variant. We also report experimental results to compare our self-dual DeepBKZ with primal BKZ and DeepBKZ for several random lattice bases.