Tiny Time Mixers

Tiny Time Mixers (TTMs) are IBM Research's compact pretrained models for multivariate time-series forecasting, introduced as the first genuinely tiny pretrained forecasters. With fewer than one million parameters, TTMs beat popular benchmarks demanding billions of parameters on zero-shot and few-shot forecasting, and they fine-tune to competitive accuracy with as little as 5% of the target training data.

Released under Apache 2.0 as part of the Granite family, TTM ships in R1, R2, and R2.1 generations trained on progressively larger public corpora (250M to roughly 1B samples), with R2.1 adding daily and weekly resolutions. Zero-shot, fine-tuning, and inference all run on a single GPU or a laptop CPU. Limitations: the open releases target point forecasting rather than probabilistic outputs, and resolution support varies by generation, so check the model card for your frequency.