Tabby

A rare fully open scientific foundation model from Huawei's Noah's Ark Lab: a 145M-parameter encoder-only patch Transformer for long-context probabilistic time series, with a context of up to 8,192 observations, released together with the complete recipe that built it. The contribution is in the data and training procedure: a corpus combining GIFT-Eval-Pretrain+ and BLAST with synthetic series from KernelSynth and CauKerV2, an online generator that composes dynamics through randomly sampled structural causal models, plus a progressive convergence schedule that yields reusable intermediate checkpoints and a deep quantile objective. It covers forecasting, classification, and anomaly detection.