TTM-R3

TTM-R3 is IBM Research's March 31, 2026 refresh of the TinyTimeMixer family and the most capable rung of the R1 → R2 → R3 ladder. While still classified as `tinytimemixer` under the hood, it moves the line from a point forecaster toward a probabilistic, structure-aware one, and it does so without abandoning the compact footprint that defines TTM. IBM reports a 15–50x inference speedup vs state-of-the-art forecasters alongside a meaningful accuracy improvement over TTM-R2.

Architecturally, R3 adds trend-residual decomposition, a multi-quantile probabilistic forecasting head, gated attention, FFT-based frequency embeddings, and learnable sequence-level register tokens, while preserving the compact ~1.4M (Lite) to ~35M parameter footprint that keeps TTM practical on CPUs and lightweight hosted inference. It is trained on IBM's GiftEvalPretrain subset plus KernelSynth-style synthetic augmentation, a corpus aligned with the broader TTM family rather than narrowed to a single benchmark. One caveat to plan around: the current upstream checkpoint and toolkit path still emit uninitialized-weight warnings during direct load, so short deterministic continuation quality should be validated on your own holdout before production reliance.

On TSFM.ai reach for R3 when you need quantile and probabilistic outputs from the tiny TTM design, or want the decomposition-aware accuracy gains over R2 — but mind the license. R3 is hosted under a pass-through compute posture and ships under CC-BY-NC-SA-4.0 (research / non-commercial use only), so you are responsible for ensuring your intended use falls within the upstream license — see section 7 of our Terms of Service. For commercial point forecasting at the same tiny footprint, stay on the Apache-2.0-licensed TTM-R2 instead.