TimesFM 3.0

The third generation of Google Research's TimesFM time-series foundation model line (original decoder-only model: ICML 2024, arXiv 2310.10688; the GitHub repo has ~28.6K stars). TimesFM 3.0 (v3.0.0, August 28, 2026) moves to a new Stacked Mixing Transformer with Variate Attention and CPM iterative RevIN: 20 transformer layers, model dim 1280, 16 heads, ~331M parameters, 32-point context patches and 64-point forecast patches with quantile outputs. Pretraining draws on GiftEvalPretrain (decontaminated against fev-bench), Wikipedia pageviews, and Google Trends data.

TimesFM is among the most deployed scientific foundation models — it backs forecasting in BigQuery ML, Google Sheets' connected-sheet forecasts, and Vertex Model Garden. Note the license change: the code remains Apache 2.0, but the 3.0 weights ship under the TimesFM Non-Commercial License v1.0, unlike earlier open checkpoints (up to 2.5) in the HF collection.