Expand description
Moirai-1.0-R-large inference engine.
Architecture: masked bidirectional transformer encoder with multi-scale patches.
- Mean-scale normalization → patch embedding (in_proj per patch size) → mask tokens for future patches → 24 encoder blocks (RMSNorm + QK-norm attention + SwiGLU FFN) → final RMSNorm → Student-t head → inverse scale → point forecast (mean = loc)