An apparatus and system for using a machine learning model for uplink beam compression are described. The model determines an excess number of streams for a UE for a target PUSCH decoding error rate. The model is trained in a near-real time (RT) radio access network (RAN) Intelligent Controller (RIC) using a decoding error observed from physical uplink shared channel (PUSCH) cyclic redundancy code (CRC) decoding of PUSCH data from the UE, in addition to the beamforming method and number of excess streams. The model is deployed in a non-RT RIC and the parameters of the deployed model is periodically updated by the near-RT RIC. Input parameters to the deployed model for the UE are provided from a distributed unit, and the output beamforming parameters provided from the deployed model to the distributed unit to provide uplink beam compression.