Abstract Purpose To compare the performance of a learned magnetization‐prepared gradient echo (L‐MPGRE) sequence against a commonly used sequence for 3D T 2 and T 1ρ mapping of the knee joint, the magnetization‐prepared angle‐modulated partitioned k‐space spoiled gradient echo snapshots (MAPSS), on bi‐exponential (BE), stretched‐exponential (SE), and mono‐exponential (ME) relaxation models. Methods We used a combined differentiable and non‐differentiable optimization to learn pulse sequence structure and its parameters for 3D T 2 and T 1ρ mapping of the knee joint using ME, SE, and BE models. The learned pulse sequence framework was used to improve quantitative accuracy and SNR and to reduce filtering effects. We compare the measured multi‐compartment values between the two sequences ( n = 8), and their repeatability ( n = 4) in healthy volunteers ( n = 12 total). Results The voxel‐wise median absolute percentage difference (MAPD) between the T 2 and T 1ρ maps obtained with each sequence was 18.6% and 19.9%, respectively. The T 2 and T 1ρ repeatability tests showed a MAPD of 18.5% and 19.1% for MAPSS, and 16.8% and 15.5% for L‐MPGRE. Bland–Altman region of interest (ROI)‐wise analysis shows that bias is small, close to −1.5%, and the coefficient of variation is around 5.5% when comparing ROIs from both sequences. Conclusion The L‐MPGRE sequences can be used as a replacement for MAPSS for T 2 and T 1ρ mapping in the knee cartilage with advantages, achieving similar accuracy and 15% better repeatability in only half of its scan time.