Carnegie Mellon University logo

MOMENT-Small

Model directory
AutonLab/MOMENT-1-small

~40M params | MIT | Upstream checkpoint

MOMENT-Small is the lightweight checkpoint in AutonLab's general-purpose time-series foundation-model family. Like the larger variants, it is framed as a multi-task representation model that transfers across forecasting, classification, anomaly detection, imputation, and embedding extraction rather than being tuned for a single task — and at the smallest size it is the cheapest way to put that backbone to work.

Architecturally it is a patch-based encoder-only transformer trained with masked time-series modeling: history is split into patches, masked spans are reconstructed during pretraining, and the resulting encoder produces representations reusable across downstream tasks. It is trained on Timeseries-PILE, built from public forecasting, classification, and anomaly-detection corpora including Informer datasets, Monash, UCR/UEA, and TSB-UAD. The upstream momentfm package warns that only the reconstruction head is pretrained and that forecasting heads must be fine-tuned, so the small variant should be treated as a compact transferable backbone rather than a plug-and-play zero-shot forecast specialist.

Where to run this model

Compute availability, hardware requirements, and pricing depend on the execution provider.

Hugging Face · Weights & deployment

Get model weights, deployment instructions, and current provider listings.

Hugging Face provider mappings checked 2026-10-06. No Inference Provider mapping was listed for this checkpoint.

Model Classification

Family

MOMENT-1

Type

time series foundation model

Pretrained time-series model for zero-shot or few-shot forecasting workloads.

Training Data

Timeseries-PILE, built from public forecasting, classification, and anomaly-detection corpora including Informer datasets, Monash, UCR/UEA, and TSB-UAD.

Recommended For

  • • Shared backbones across forecasting, anomaly detection, classification, and imputation
  • • Teams that want one general-purpose time-series representation model

Strengths

  • • Multi-task time-series understanding
  • • Useful when the same deployment needs to cover several downstream tasks

Limitations

  • • Not optimized purely around one forecasting leaderboard objective
  • • May be heavier than needed if you only need straightforward zero-shot forecasting
  • • The upstream momentfm package warns that only the reconstruction head is pretrained and that forecasting heads must be fine-tuned
  • • Hosted forecasting quality can lag specialist zero-shot forecasters because MOMENT is primarily framed as a transferable representation backbone
  • • Zero-shot forecasting tends to be trend-blind — predictions may flatten regardless of clear trends in the input

Not Ideal For

  • • Choosing a default forecast-only model when you mostly care about zero-shot continuation quality
  • • Short benchmark-style trend extrapolation where specialist forecasting families are stronger

Capabilities

forecastingclassificationanomaly-detectionimputationretrieval

Tags

momentmulti-taskrepresentation-learninglightweight

Specifications

Parameters
~40M
Architecture
patch-based encoder-only transformer trained with masked time-series modeling
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
MIT

Related models