Marchuk-S2S: Efficient Latent Flow Matching for Subseasonal Ensemble Weather Forecasting

Dense six-hourly ensemble forecasts capture subdaily weather evolution and forecast uncertainty that daily averages can obscure, offering potential value for emergency response. We introduce Marchuk-S2S, a compact 276M-parameter autoregressive latent flow-matching Transformer that efficiently generates six-hourly ensembles across medium-range and subseasonal horizons. The model combines variable-horizon training, timestamp-based annual-cycle conditioning, learned positional embeddings, and CRPS fine-tuning. When aggregated to daily resolution, Marchuk-S2S retains forecast skill competitive with FuXi-S2S and ECMWF S2S. Initial evaluations indicate stable behavior over 48-day rollouts, providing preliminary evidence of long-horizon stability.