FlowState r1.1

FlowState r1.1 is IBM's March 31, 2026 refresh of the sampling-rate-invariant FlowState family and the recommended default over the original r1. It keeps the sampling-rate-invariance and variable-horizon properties that defined r1 while scaling up the model behind them. According to IBM, r1.1 currently leads the GIFT-Eval leaderboard for point-forecasting accuracy by MASE.

Compared with r1 it adds an output-gating mechanism to the S5 state-space encoder, extends the pretraining context from 2048 to 4096 time points, and enlarges the MLP, roughly doubling the parameter count from ~9M to ~18.5M while preserving the functional basis decoder. Pretraining also adds CauKer-generated synthetic series on top of the r1 corpus of Gift-Eval Pretrain and Chronos pretraining subsets; IBM states none of the training data overlaps the Gift-Eval evaluation splits. The r1.1 weights are published on the ibm-research/flowstate repository under the r1.1 revision branch.

On TSFM.ai reach for r1.1 when you want IBM's strongest point-forecasting accuracy combined with FlowState's cross-cadence flexibility, particularly on longer-context series that benefit from the extended 4096-point window. It is hosted alongside r1 so existing workloads can A/B test before migrating; stay on r1 only if you need to reproduce a prior result on the smaller checkpoint.