Foundation ModelTHUML, Tsinghua University

Sundial

Generative TSFM pretrained on TimeBench (~1 trillion time points). Introduces TimeFlow Loss to predict next-patch distributions directly, removing the need for discrete tokenisation and enabling non-deterministic, probabilistic forecasts. ICML 2025 Oral.

Sundial is a generative time-series foundation model from Tsinghua's THUML group. Where Chronos quantises values into a discrete vocabulary, Sundial keeps continuous outputs and trains with a novel “TimeFlow Loss” that directly predicts the distribution over the next patch. This lets the transformer emit non-deterministic, probabilistic forecasts without ever tokenising.

Pretraining uses TimeBench — about one trillion time points, predominantly real-world data with a synthetic component. Sundial reports SOTA results on the Time-Series-Library benchmark, GIFT-Eval, and FEV, and the released 128M checkpoint runs zero-shot inference on CPU within seconds.

Versions on TS-Arena

Each version below corresponds to one registered model id in the leaderboard. Click through to its detail page for per-model rankings, forecasts, and history.

  • Sundial Base 128M
    sundial-128m
    128M params

    Single released checkpoint at the time of writing; pretrained on TimeBench.