Flowstate, Tiny Time Mixer (TTM), and Time Series Pulse (TSPulse) are IBM’s family of
ultra-lightweight pre-trained models for time series data, with just a few
million parameters and GPU-free inference. Despite their compact size, these
models deliver strong performance across forecasting, classification, anomaly
detection, imputation, and similarity search tasks.FlowState is a time-scale adjustable foundation model, designed to offer flexible and efficient forecasting across various temporal scales. By integrating a State Space Model (SSM) Encoder with a Functional Basis Decoder, FlowState transitions into a timescale-invariant coefficient space, enabling continuous forecasting adaptable to any sampling rate. This architecture allows for training at one timescale and inference at another, significantly enhancing the utility of training data across diverse temporal structures. FlowState delivers state-of-the-art accuracy in zero-shot forecasting tasks, holding the #3 spot on the GIFT-Eval leaderboard (as of 10/2/2025).TTM models supports multivariate forecasting via both channel independence and
channel-mixing approaches. Decoder Channel-Mixing can be enabled during
fine-tuning for capturing strong channel-correlation patterns across time-series
variates, a critical capability lacking in other existing approaches. In
addition, TTM also supports exogenous and categorical data infusion.TSPulse introduces a novel dual-space masked reconstruction that jointly learns
from time and frequency domains, unifying complementary signals into a shared
embedding space. A dual-embedding disentanglement mechanism separates
fine-grained and semantic features, while task-relevant heads are prioritized
during pretraining through reweighted loss objectives. This produces
task-specific refined representations, enabling efficient transfer learning for
downstream tasks.
