Granite-TimeSeries-TSPulse-R1

TSPulse is IBM's ultra-compact time-series understanding model, built for anomaly detection, classification, imputation, and similarity search rather than forecasting. It is the odd one out in the Granite TSFM lineup: instead of predicting the future, it characterizes and repairs the series you already have, and it does so at a footprint of around 1M parameters.

Architecturally it is an MLP-Mixer that performs dual-space (time + frequency) masked reconstruction, which is what gives it strong understanding behavior at such a small size. The official card reports +20% on TSB-AD, +25% on semantic similarity search, and +50% in zero-shot imputation. IBM's public card emphasizes broad multi-domain time-series pretraining for these tasks but does not publish a detailed source-by-source corpus breakdown.

TSPulse is tracked on TSFM.ai as part of the Granite TSFM family but is not currently hosted on our inference runtime, because it is not a forecasting model and our API surface is forecasting-centric. When you need anomaly detection, classification, imputation, or similarity search rather than a forecast, this is the right model in the family — but run it from the Hugging Face repository directly. For forecasting tasks, use one of the hosted Granite checkpoints such as TTM or FlowState instead.