FlowState

FlowState is IBM's sampling-rate-invariant TSFM for zero-shot forecasting, and the original release in the FlowState family. Its defining trait is timescale flexibility: rather than being tied to a single fixed cadence, one model is meant to generalize across temporal resolutions, which is what sets it apart from the fixed-timescale checkpoints elsewhere in the catalog.

Architecturally it combines a state-space (SSM) encoder with a functional basis decoder, a pairing that lets it adapt context length, target length, and sampling rate at inference time instead of baking those in at training. At 9.07M parameters it is compact. It is pretrained on subsets of Gift-Eval Pretrain and the Chronos pretraining corpus, and the official card states that the data used does not overlap the Gift-Eval evaluation splits.

On TSFM.ai reach for FlowState when your data arrives at inconsistent cadences, or when you need one model to span multiple temporal resolutions without retraining per frequency. For new workloads, prefer the FlowState r1.1 refresh — it keeps these same sampling-rate-invariance and variable-horizon properties with a larger context and capacity; r1 remains hosted alongside it so you can A/B test before migrating.