Time-MoE-50M

Time-MoE-50M is the smaller public checkpoint in Xiaohongshu's Time-MoE family. The family naming refers to its active-parameter regime: this checkpoint activates roughly 50M parameters per forward pass even though it stores about 0.1B in total, because the inactive experts remain part of the release. It is the lightest public way to evaluate the sparse Time-MoE design.

Architecturally it is a decoder-only transformer with sparse Mixture-of-Experts routing, which lets it reach for large-capacity behavior without paying the full dense compute cost on every token. It is pretrained on Time-300B, the multi-domain corpus spanning more than nine domains and over 300B time points described in the official paper and repository, and it is released under Apache-2.0.

On TSFM.ai reach for Time-MoE-50M when you want to explore MoE forecasting behavior at the lowest cost, or as a cost-efficient high-throughput baseline. Step up to Time-MoE-200M when you want more capacity and longer-context autoregressive forecasting from the same sparse-expert design; the larger checkpoint activates roughly 200M parameters per pass against about 0.5B stored.