🚀 Update (Mar 2026): TSPulse has been accepted at ICLR 2026.
TSPulse models are ultra-compact pre-trained models for time-series data, featuring just 1M parameters and supporting GPU-free inference. Designed for versatility, TSPulse excels across a wide range of tasks including classification, anomaly detection (AD), imputation, and similarity search.
At the architecture level, TSPulse introduces a novel dual-space masked reconstruction strategy that learns jointly from both time and frequency domains —- capturing complementary patterns in a unified embedding space. This is coupled with a dual-embedding disentanglement mechanism that produces both fine-grained embeddings for detailed analysis and semantic embeddings for broader understanding. These semantic embeddings are inherently robust to variations in time, magnitude, and noise, making them ideal for time-series retrieval.
At the task level, TSPulse integrates several innovations:
TSLens: A fine-tuning module for task-aware feature extraction.
Multi-head triangulation: Fuses outputs from multiple prediction streams to enhance anomaly detection robustness.
Hybrid masking: Reduces bias during pre-training, improving zero-shot imputation.
TSPulse demonstrates significant improvements on several benchmarks:
+5–16% on UEA classification benchmarks
+20% on the TSB-AD benchmark for anomaly detection (rank #1 in both univariate and multivariate)
+50% in zero-shot imputation
+25% in semantic similarity search
Despite being 10–100× smaller than competing models, TSPulse sets a new standard in efficient time-series foundation modeling—combining strong out-of-the-box performance, minimal compute requirements, and rapid adaptability to downstream tasks.
Inspired by the success of small, task-specialized pre-trained models in the language/vision domain which achieve strong performance through minimal task-specific adaptations —- we extend this strategy to time-series. Since TSPulse offers multiple heads we specialize the pre-training for every task through reweighting loss objectives to prioritize heads most relevant to the target task. This enables TSPulse to refine task-specific representations while maintaining its lightweight design, facilitating efficient transfer learning across any datasets for the specified downstream task.
from tsfm_public.models.tspulse import TSPulseForReconstruction
# ✅ For Imputation and Semantic Search
model = TSPulseForReconstruction.from_pretrained(
"ibm-granite/granite-timeseries-tspulse-r1",
revision="tspulse-hybrid-dualhead-512-p8-r1"
)
# ✅ For Anomaly Detection
model = TSPulseForReconstruction.from_pretrained(
"ibm-granite/granite-timeseries-tspulse-r1",
revision="main"
)
# ✅ For Classification
from tsfm_public.models.tspulse import TSPulseForClassification
model = TSPulseForClassification.from_pretrained(
"ibm-granite/granite-timeseries-tspulse-r1",
revision="tspulse-block-dualhead-512-p16-r1"
)
The model is trained with a base context length of 512. Different tasks have varying minimum input length requirements:
Anomaly Detection (AD): Requires time-series of at least 3–4× the base context length (i.e., ~1536–2048 points) to ensure stable and robust anomaly detection.
Imputation & Semantic Search: Require a minimum of 512 points to align with the model's context window and enable meaningful processing.
Classification: Supports any input length. Internally, the model resamples or interpolates the input to a fixed length of 512 before inference.
Please cite the following paper if you intend to use our model or its associated architectures/approaches in your work.
BibTeX:
@misc{ekambaram2025tspulsedualspacetiny,
title={TSPulse: Dual Space Tiny Pre-Trained Models for Rapid Time-Series Analysis},
author={Vijay Ekambaram and Subodh Kumar and Arindam Jati and Sumanta Mukherjee and Tomoya Sakai and Pankaj Dayama and Wesley M. Gifford and Jayant Kalagnanam},
year={2025},
eprint={2505.13033},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.13033},
}
All content in this repository including code has been provided by IBM under the associated open source software license and IBM is under no obligation to provide enhancements, updates, or support. IBM developers produced this code as an open source project (not as an IBM product), and IBM makes no assertions as to the level of quality nor security, and will not be maintaining this code going forward.
Files
5 files · 4.68 MBWeights 4.37 MB (safetensors). Sizes and sha256 from the signed manifest.
