Google Research released TimesFM 3.0, a pretrained model for forecasting time-based data. The new version forecasts several related series at once, rather than one at a time. It also accepts extra input series, including ones with known future values. Google says no per-task tuning is needed for this. Google reports first place on three forecasting benchmarks: fev-bench, the TIME Benchmark and GIFT-Eval. The 3.0 weights carry a non-commercial licence and may not be used in production. Earlier weights up to version 2.5 stay under the permissive Apache-2.0 terms. The code itself remains Apache-2.0. TimesFM is also offered inside BigQuery ML, Google Sheets and Vertex Model Garden. A blog post explaining 3.0 has not been published yet.
What changed
TimesFM 2.5 forecast one series at a time, with covariates only through an add-on, and its weights were Apache-2.0.
What it unlocks
Forecasting several related series jointly, with known future inputs such as promotions or holidays, without task-specific training.
- rank #1 on fev-bench, 100 tasks
- TIME Benchmark: 50 datasets, 98 tasks
- 2.5 model: 200M parameters, 16k context
What you need to act on it
- accepting the non-commercial licence for the 3.0 weights
- a GPU and Python setup, or use via BigQuery ML, Google Sheets or Vertex Model Garden
Sources