Granite-TimeSeries-TTM-R2

Granite-TimeSeries-TTM-R2 is IBM's larger-data continuation of the TinyTimeMixer line and the natural step up from R1. IBM positions it as a better-performing follow-on to R1 while preserving the small, fast deployment profile that makes TinyTimeMixer practical on CPUs and lightweight hosted inference. Like R1 it remains a focused family of context- and horizon-specific checkpoints rather than a single universal TSFM, so accuracy comes from matching a checkpoint to your series rather than from raw scale.

At 805K parameters R2 keeps the same compact TinyTimeMixer footprint as R1; the gain comes from training data rather than capacity. Its public forecasting corpus is built from the same domains as R1 — Australian Electricity and Weather, Bitcoin, KDD Cup 2018, London Smart Meters, Saugeen River Flow, Solar, US Births, and wind datasets — but IBM states R2 used about 700M training samples, roughly tripling R1's volume, which is what underpins the accuracy improvement at an unchanged serving cost.

On TSFM.ai pick R2 as the default tiny IBM forecaster: it is the better-performing checkpoint at the same throughput and deployment profile as R1, so there is little reason to stay on R1 unless you are reproducing a prior result. Move to TTM-R3 when you need probabilistic quantile outputs and the decomposition-aware refresh of the family, noting its different license.