Sundial

Summary

Sundial is THUML’s flow-matching time-series forecasting foundation-model family. It anchors continuous probabilistic forecast heads over numeric patches, while remaining a passive forecasting model with univariate pretraining and no explicit covariate, action, control-input, or intervention channel.

Role In The Wiki

Sundial is useful for comparing continuous flow-matching forecast heads against discrete-token, quantile-head, and deterministic forecasting interfaces. It also sharpens the scaling question because its reported gains mix TimeFlow objective design, TimeBench corpus scale, and Transformer/inference engineering.

Official Artifacts

• Official code: https://github.com/thuml/Sundial
• Official checkpoint: https://huggingface.co/thuml/sundial-base-128m

Evidence

• Sundial source page

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in sundial-2025 rather than duplicating verdict rows here.

At the entity level, Sundial is useful for comparing continuous flow-matching forecast heads against discrete-token, quantile-head, and deterministic forecasting interfaces. It also sharpens the scaling question because its reported gains mix TimeFlow objective design, TimeBench corpus scale, and Transformer/inference engineering. This page should stay as the object card; source pages carry slot-level verdicts, evidence, and missing pieces.

Related Pages

• Foundation Time-Series Model Research Agenda
• Time-Series Foundation Models
• Synthetic Data For Time Series
• Time-Series Scaling And Efficiency
• Time-Series Benchmark Hygiene
• High-Dimensional Time Series Forecasting
• T2S