Moirai-1.0-R-Base

Moirai-1.0-R-Base is the reference dense checkpoint in Salesforce's original Moirai family, sitting between Moirai-1.0-R-Small and Moirai-1.0-R-Large. At 91M parameters it is the natural default for the first-generation Moirai design: enough capacity to improve general forecasting quality on heterogeneous multivariate settings without the cost of the large variant.

It keeps the same masked-encoder transformer used across the dense 1.0-R line, with multi-patch projections, any-variate attention, and a mixture-distribution output head. The any-variate attention lets a single model handle arbitrary numbers of target variables and dynamic covariates, while the mixture output head returns probabilistic forecasts with calibrated quantiles rather than a single trajectory. Like its siblings it is pretrained on the LOTSA corpus, giving it broad zero-shot coverage across domains.