t0-alpha Shows Small Transformer Can Match Large Time‑Series Forecasting Benchmarks
t0-alpha, a 102M-parameter decoder-style transformer, demonstrates how modern time-series foundation models tokenise, attend causally, and output probabilistic quantiles. Running it reproduces the paper's CRPS 0.4941 and MASE 0.7240, showing competitive accuracy while remaining hardware‑friendly. Its patch‑based design points to calibration and routing as the next big leaps.