Moirai-2.0-R-Small

Moirai-2.0-R-Small is the faster successor to the original dense Moirai family, the small checkpoint in the newer 2.0-R line. At roughly 11M parameters it is a deliberately compact model: Salesforce reports performance that surpasses larger earlier Moirai checkpoints, so the 2.0 redesign is meant to deliver more accuracy per parameter than the first-generation 1.0-R and 1.1-R families.

It marks an architectural break from the masked-encoder Moirai design, switching to a decoder-only transformer with quantile loss and multi-token prediction, plus better missing-value handling. The quantile loss yields calibrated probabilistic multivariate forecasts directly, while multi-token prediction speeds up decoding. Pretraining draws on a subset of GIFT-Eval Pretrain and Train, Chronos mixup data, KernelSynth synthetic series, and internal Salesforce operational data, as listed in the official model card.

On TSFM.ai it is served at the published 512-step configuration for this family rather than advertising longer unsupported histories. Reach for it when you want the most modern, efficient Moirai checkpoint with strong zero-shot accuracy at a small footprint; choose a dense 1.0-R or 1.1-R Moirai variant instead when you specifically need the earlier masked-encoder architecture or its explicit covariate support.