Marchuk: A Compact Latent Flow-Matching Model for Stochastic Weather Forecasting

We present Marchuk, a compact 276M-parameter latent flow-matching Transformer for probabilistic global weather forecasting from medium-range to subseasonal scales. Conditioned on one atmospheric state and calendar phase, Marchuk jointly samples dense 6-hourly trajectories, is trained on 1-, 2-, 4-, and 8-day chunks, and autoregressively composes four-day chunks to 45 days with one shared network. The model combines variable-horizon training, separated spatial and temporal conditioning, and latent-space CRPS fine-tuning, and remains competitive with larger latent and operational ensemble baselines while enabling efficient 50-member forecasts. Weekly anomaly-correlation and long-rollout structural diagnostics on held-out 2018 forecasts examine retained subseasonal signal and spatial variability. A direct-versus-composed transition study finds narrower ensembles with comparable spread and improved RMSE and CRPS under shorter compositions, indicating partial compatibility of endpoint marginals without a consistency loss.