Sundial (Base 128M)
Model directorythuml/sundial-base-128m128M params | Upstream checkpoint
Sundial is THUML's native generative TSFM for continuous-valued forecasting. The official card presents it as a 128M-parameter model built to generate multiple plausible futures rather than a single trajectory, making it one of the most directly probabilistic models in the model landscape and a strong option when distribution quality matters as much as point accuracy.
Architecturally it is a causal decoder-only transformer trained with a TimeFlow / flow-matching loss, which lets it model continuous values natively instead of relying on the discrete tokenization used by many other forecasting families. It is pretrained on TimeBench, a corpus of 1032B time points, with the model card listing THUML's UTSD, LOTSA, and the Chronos datasets as core upstream sources.
Where to run this model
Compute availability, hardware requirements, and pricing depend on the execution provider.
Get model weights, deployment instructions, and current provider listings.
- Community forecasting playground
hosted demo · Source checked 2026-10-06
The published loader names this checkpoint. Space was sleeping when checked; it may require a wake-up. Hosted computation, not browser inference.
Hugging Face provider mappings checked 2026-10-06. No Inference Provider mapping was listed for this checkpoint.
Model Classification
Family
Sundial
Type
time series foundation model
Pretrained time-series model for zero-shot or few-shot forecasting workloads.
Resources
Training Data
TimeBench / 1032B time points, with the model card listing THUML UTSD, LOTSA, and Chronos datasets as core upstream sources.
Recommended For
- • Probabilistic forecasting where distribution quality matters
- • Teams that want native continuous generative forecasting rather than tokenization
Strengths
- • Designed around probabilistic future generation
- • Native continuous-valued modeling rather than discrete token bins
Limitations
- • Less suited to broad non-forecasting tasks
- • Generative probabilistic focus can be overkill for simple point-forecast-only use cases
Capabilities
Tags
Specifications
- Parameters
- 128M
- Architecture
- causal decoder-only transformer with TimeFlow loss / flow matching
- Context length
- n/a
- Max context
- n/a
- Minimum history
- n/a
- Recommended history
- n/a
- Input step
- n/a
- Required target series
- 1
- Temperature
- Ignored
- Top P
- Ignored
- Max output
- n/a
- License
- n/a