OSS
google-research/timesfm: Google's Time Series Foundation Model v2.5 Surges to 11.7K Stars With Workspace Integration
TimesFM 2.5 adds covariate support through XReg and achieves comparable performance to v2.0 at a smaller footprint via QKV matrix fusion. Google integrated TimesFM into Connected Sheets (rollout started Feb 2026), enabling business users to generate forecasts without ML pipelines. The model uses 50 transformer layers at 1,280-dimensional width with 4x longer context than v1.0 and trains on LOTSA archive including cloud infrastructure traces, solar/wind, and climate data.
Source
↳ Follow the thread