Objects in hyperspectral images (HSI) exist many subtle information differences, thus multi-level spectral-spatial perception will be beneficial to discriminative feature learning for HSI. We propose a multi-level mixed perception network (MMPN) for HSI classification, which is composed of three perceptrons: compact global and partition spectral perceptron (CSeP), pixel-wise spectral-partition perceptron (PSeP), and local spatial perceptron (LSaP). Specifically, we partition the object-centered block from HSI into non-overlapping spectral patches equidistantly. CSeP is designed on the squeezed feature to model spectral dependencies from overall and intra patches, respectively. The outputs are embedded together into the original patches for spectral information calibration. Then, PSeP is followed to avoid subtle spectra confusion, and LSaP is concurrently followed for multiscale spatial feature extraction. The learned features from each patch are used for label prediction respectively, and finally soft voting the classification result. Experimental results across two HSI datasets indicate that MMPN achieves expect performance in object classification when compared with the state-of-the-art methods.