Abstract Deep learning has revolutionized short‐term Sea Surface Temperature (SST) forecasting, but often struggles with cumulative error propagation and operational complexity in medium‐range predictions. To achieve lightweight and temporally continuous SST forecast, this study introduces two novel training strategies, Rolling Training and Multistep Training, to optimize autoregressive forecasting within an Earthformer neural network. By enforcing a consistent 1‐day rolling inference process, our framework eliminates temporal discontinuities and forecast ambiguity inherent in traditional Sequence‐to‐Sequence (Seq2Seq) models. We validate this optimized model, named Multi‐Earthformer, using OISST data and verify its robustness against independent in situ Argo observations. Results demonstrate that Multi‐Earthformer significantly enhances predictive skill beyond 15‐day lead times, suppressing seasonal systematic biases and maintaining stable, quasi‐linear error growth. Our approach offers a lightweight, physically intuitive paradigm that bridges the gap between daily synoptic variability and subseasonal‐to‐seasonal evolution, providing a scalable solution for reliable marine forecasting without the computational burden of traditional multi‐model ensembles.