FlowState is a new forecasting model designed to work across data sampling rates without being retrained for each one.
Researchers published FlowState, a time-series foundation model that pairs a state space model encoder with a functional basis decoder. Together, the design lets the model operate in continuous time rather than discrete steps, so it handles hourly sensor data, daily financial feeds, and irregular IoT signals with the same weights. The model also adjusts forecasting horizons on the fly. Despite being smaller than most competing time-series foundation models, FlowState topped the GIFT-Eval benchmark, a standard test for generalist forecasting accuracy.
The sampling-rate problem is easy to underestimate until you try to deploy a forecasting model in production. Real data pipelines mix rates constantly: a factory floor might log temperature every 200 milliseconds and energy consumption every 15 minutes. Retraining a separate model for each resolution is expensive; stitching together workarounds is fragile. A model that genuinely handles arbitrary rates without that overhead addresses a pain point that most benchmark papers skip over entirely.
State space models have been promising transformer-level performance at lower compute for a while now — the real question is whether this benchmark lead holds up once someone tries it on their own messy data.