In IMDD systems, the nonlinear impairment caused by the interaction between the transceiver and the fiber dispersion seriously reduces the signal-to-noise ratio (SNR). Piecewise linear (PWL) equalizers have lower complexity and higher stability in nonlinear compensation compared to Volterra, but it has difficulty in adaptively partitioning due to the gradient loss caused by the hard partition. We propose a novel adaptive PWL equalizer using a soft partition mechanism, named soft-PWL (SPWL). The soft partition uses multiple binary classifiers to computing partition probability with gradient. Thus, we can realize the adaptive partition by the method of gradient descent. This greatly improves the performance of equalization, and simultaneously enables its flexibly application in nonlinear dominant scenarios. We conducted an experiment of 80-Gb/s PAM-6 signal transmission over 20-km single mode fiber. The experimental results show that the SPWL is more robust to nonlinearity strength than Volterra/DNN. The computational complexity (CC) of SPWL is about 41% lower than that of Volterra at the same bit error rate (BER). The SPWL realizes the adaptive partition and improves receiver sensitivity by 3 dB compared to conventional PWL with an SD-FEC threshold of 2.2 × 10 −2 .