Moirai

Moirai, from Salesforce AI Research (arXiv February 2024, ICML 2024 oral), is a universal time-series forecasting transformer trained on LOTSA, a 27-billion-observation corpus spanning nine domains. The masked-encoder architecture handles arbitrary frequencies, variate counts, and forecast horizons in a single model, shipping in Small (36M), Base (94M), and Large (305M) sizes with a later Mixture-of-Experts variant and a decoder-only Moirai 2.0. It was the first fully open-source time-series foundation model — weights, data, and code all public via the uni2ts library.

Limitations: independent Nixtla benchmarks found classical statistical ensembles often still beat Moirai on accuracy for standard frequencies, with Moirai 160x slower than seasonal-naive baselines — the classic foundation-model trade-off. Moirai 2.0 narrowed the gap (2x faster, 30x smaller than 1.0-Large) but long-horizon forecasting remains a weakness.