Time-MoE-200M

Time-MoE-200M is the larger public checkpoint in Xiaohongshu's Time-MoE family. The official paper and repo position the family as a way to scale time-series model capacity more efficiently than dense decoder-only models while keeping long-context autoregressive forecasting practical, and this checkpoint is the higher-capacity option in that line.

Architecturally it is a decoder-only transformer with sparse Mixture-of-Experts routing: the published checkpoint reports about 0.5B stored parameters while activating only roughly 200M per forward pass, so it buys large-model capacity without dense-model inference cost. It is pretrained on Time-300B, the multi-domain corpus spanning more than nine domains and over 300B time points described in the official materials, and it is released under Apache-2.0.

On TSFM.ai reach for Time-MoE-200M when you want the strongest Time-MoE behavior, particularly for long-context autoregressive forecasting where the extra capacity pays off. Drop down to Time-MoE-50M when you want the same sparse-expert design at the lowest cost and highest throughput rather than peak accuracy.