OBJECTIVES: Humanised mice have emerged as valuable models for pre-clinical testing of the safety and efficacy of immunotherapies. Given the variety of models available, selection of the most appropriate humanised mouse model is critical in study design. Here, we aimed to develop a model for predicting cytokine release syndrome (CRS) while minimising graft-versus-host disease (GvHD). METHODS: To overcome donor-induced variation, we directly compared the in vitro and in vivo immune phenotype of immunodeficient NSG mice reconstituted with human bone marrow (BM) CD34+ haematopoietic stem cells (HSCs), peripheral blood mononuclear cells (PBMCs) or spleen mononuclear cells (SPMCs) from the same human donors. SPMC engraftment in NSG-dKO mice, which lack MHC class I and II, was also evaluated as a strategy to limit GvHD. Another group of mice was engrafted with umbilical cord blood (UCB) CD34+ HSCs. Induction of CRS in vivo was investigated upon administration of the anti-CD3 monoclonal antibody OKT3. RESULTS: PBMC- and SPMC-reconstituted NSG mice showed short-term survival, with engrafted human T cells exhibiting mostly an effector memory phenotype. Survival in SPMC-reconstituted NSG-dKO mice was significantly longer. Conversely, both BM and UCB-HSC models showed longer survival, without demonstrable GvHD and a more naïve T-cell phenotype. PBMC- and SPMC-reconstituted mice, but not BM-HSC or UCB-HSC mice, experienced severe clinical signs of CRS upon administration of OKT3. CONCLUSION: PBMC- and SPMC-reconstituted NSG mice better predict OKT3-mediated CRS. The SPMC model allows generation of large experimental groups, and the use of NSG-dKO mice mitigates the limitation of early GvHD.