Foundation ModelSalesforce AI Research

Moirai

Masked-encoder universal forecasting transformer trained on LOTSA (~27B observations across nine domains). Multiple patch-size projection layers handle frequency diversity; an any-variate attention mechanism handles arbitrary numbers of covariates; a mixture-distribution head models flexible predictive distributions.

Moirai is Salesforce's “universal” forecasting transformer. The 1.x line is a masked-encoder model: history patches are encoded jointly, masked target patches are reconstructed, and the same model handles univariate and multivariate inputs through an any-variate attention mechanism that ignores variate ordering.

It addresses three classic time-series challenges directly: cross-frequency learning (multiple patch-size projection heads chosen per frequency), arbitrary covariate count (any-variate attention), and varying distributional shapes (a mixture-distribution output head). Pretraining uses the LOTSA archive — ~27B observations spanning nine domains.

Moirai-2.0-R (Nov 2025) revisits the architecture as a decoder-only model and re-trains on the GIFT-Eval Pretrain subset with additional data, currently available in a Small (14M) variant.

Versions on TS-Arena

Each version below corresponds to one registered model id in the leaderboard. Click through to its detail page for per-model rankings, forecasts, and history.

  • Moirai 1.1-R Small
    moirai-small
    14M params

    Encoder-only. Smallest 1.1-R checkpoint.

  • Moirai 1.1-R Base
    moirai-base-model
    91M params

    Encoder-only.

  • Moirai 1.1-R Large
    moirai-large
    311M params

    Encoder-only. Largest 1.1-R checkpoint.

  • Moirai 2.0-R Small
    moirai-2-small
    14M params

    Second-generation Moirai. Decoder-only; pretrained on the GIFT-Eval Pretrain subset plus additional data.