Sundial

Sundial (Liu et al., ICML 2025) is Microsoft Research's entry in the time-series foundation model race, and its distinguishing idea is how it handles uncertainty. Instead of quantizing values like Chronos, it keeps the series continuous: patch embeddings feed a decoder-only transformer with RoPE, and a parameterized TimeFlow Loss models the per-token distribution and generates raw values directly at inference. The family includes 128M and larger variants, supports context up to 2,880 steps, and does probabilistic forecasting, anomaly detection, and imputation out of the box.

The authors report best-in-class zero-shot scores on their evaluation: MASE 0.673 and CRPS 0.472, ahead of Chronos, Moirai, and TimesFM on the FEV and GIFT-Eval suites they ran. Checkpoints are published on Hugging Face (thuml/sundial-base-128m), though no license terms were published for them at the time of research. As a 2025 model it has less community validation than Chronos or Moirai, and the reported wins are provider-run, so treat the leaderboard as provisional.

Author-reported (Liu et al., ICML 2025, arXiv 2502.00816): best MASE (0.673) and CRPS (0.472) among zero-shot foundation models evaluated (Chronos, Moirai, TimesFM, TabPFN) on FEV and GIFT-Eval; independent verification is limited so far given the 2025 release date.