The Forecasting Company released t0-alpha, a 102M-parameter decoder-style transformer for probabilistic time-series forecasting. This model breaks down raw series into 32-step patches, processes them using causal time-attention, and predicts future quantiles. Its open weights and manageable size make it ideal for understanding the architecture behind modern time-series foundation models.

Published under Apache-2.0, t0-alpha ran its benchmark tests on GIFT-Eval, precisely reproducing reported numbers: CRPS 0.4941 and MASE 0.7240. This probabilistic approach, outputting a distribution of potential futures rather than a single point, is a core characteristic differentiating these models from earlier forecasting methods.

Unlike models that adapt text LLMs, t0-alpha trains natively on time-series data. It exemplifies a stable recipe: a transformer backbone, broad pretraining on time-series datasets, probabilistic outputs, and zero-shot evaluation across diverse domains. This approach offers a cleaner understanding of the current generation of forecasting tools.