Mantis, a new foundation model trained entirely on synthetic data, claims a state-of-the-art result for time-series classification.
Researchers built Mantis as a transformer pre-trained via self-supervised contrastive learning, using no real-world labeled data. The team argues that tokenization is the key unlock: they developed a new "token generator unit" to slice time-series inputs before they reach the model. At test time, Mantis uses intermediate-layer representations, self-ensembling, and cross-model embedding fusion to close the gap against specialized models. Across four dataset collections spanning multiple domains, it outperformed existing foundation models.
Most foundation model research on time series has fixated on forecasting, meaning predicting what comes next, while classification has lagged. A general-purpose classifier that works across domains without domain-specific training data has clear appeal for anything from medical waveform analysis to industrial sensor monitoring.
The synthetic-data-only training is the interesting bet. It either signals genuine generalization far beyond the training distribution, or it means the current benchmarks don't yet capture real deployment conditions.