TimesFM 3.0
Google Research · USA · 2026
A 330-million-parameter forecasting model that reads several series at once and ranks first on all three major time-series benchmarks. It is also the first TimesFM whose weights may not be used commercially.
TimesFM 3.0 is the third generation of Google Research's foundation model for time-series forecasting, announced in the project repository on 28 August 2026. It forecasts without being trained on the user's data — the model sees a history of measurements and returns a horizon, zero-shot, the way a language model completes a sentence. What is new in this generation is that it is natively multivariate: earlier TimesFM models handled one series at a time, while 3.0 takes several channels together and accepts covariates, both those known only for the past and those known for the future as well. In practice that is the difference between forecasting sales from sales alone and forecasting them alongside weather, prices and a calendar of holidays. The model is small by the standards of this catalogue — 330 million parameters in 20 transformer layers of 1,280 dimensions with 16 heads, reading history in patches of 32 points and emitting a horizon in patches of 64, for up to 32 variates at once. It does not return a single number but nine quantiles from 0.1 to 0.9, so the forecast comes with its own uncertainty. Training data, as Google lists it, is the GiftEvalPretrain collection with the sets that overlap fev-bench removed, Wikipedia pageviews up to November 2023, the most frequent Google Trends queries up to the end of 2022, and synthetic and augmented series. Google reports first place on all three benchmarks that matter in this field: rank one overall on fev-bench across 100 real-world tasks, rank one overall on the TIME benchmark across 50 datasets and 98 tasks, and rank one among foundation models on GIFT-Eval. The catch is the licence. TimesFM 1.0, 2.0 and 2.5 all shipped their weights under Apache 2.0, and the source code still does. The 3.0 weights do not: they carry a separate TimesFM Non-Commercial Licence v1.0 that permits testing, evaluation and academic research, forbids production deployment, revenue-generating use, and distillation into commercial models, and forbids redistributing the weights at all. Commercial users are pointed at Google's own products, where the model is served under different terms — BigQuery ML, Connected Sheets and the Vertex AI Model Garden.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!