We consider a multi-user multiple-input multiple-output (MU-MIMO) system that uses orthogonal frequency division multiplexing (OFDM). Several receivers are developed for data detection of MU-MIMO transmissions where two users share the same OFDM time and frequency resources. The receivers have partial state information about the MU-MIMO transmission with each receiver having knowledge of the MU-MIMO channel, however the modulation constellation of the co-scheduled user is unknown. We propose a joint maximum likelihood (ML) modulation classification of the co-scheduled user and data detection receiver using the max-log-MAP approximation. It is shown that the decision metric for the modulation classification is an accumulation over a set of tones of Euclidean distance computations that are also used by the max-log-MAP detector for bit log-likelihood ratio (LLR) soft decision generation. An efficient hardware implementation emerges that exploits this commonality between the classification and detection steps and results in sharing of the hardware resources. Comparisons of the link performance of the proposed receiver to several linear receivers is demonstrated through computer simulations. It is shown that the proposed receiver offers \unit[1.5]{dB} improvement in signal-to-noise ratio (SNR) over the nulling projection receiver at $1\%$ block error rate (BLER) for $64$-QAM with turbo code rate of $1/2$ in the case of zero transmit and receiver antenna correlations. However, in the case of high antenna correlation, the linear receiver approaches suffer significant loss relative to the optimal receiver.