Google Research posts TimesFM-3: 330M parameters, multivariate one-pass, non-commercial weights

Google Research published TimesFM-3 on 31 August 2026: a 330 million parameter time-series foundation model pretrained on more than 1 trillion time points. The Google Research blog says it can jointly forecast related series, plus past covariates and known future signals such as holidays or promotions, in one forward pass, without a separate fine-tune per task.

From the 31 August 2026 @GoogleResearch post: "Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks."

Key facts: - Weights live at `google/timesfm-3.0-pytorch`. GitHub tagged TimesFM-3.0 (v3.0.0) on 28 August 2026. The official blog and X post landed on 31 August. - Decoder-only transformer with 32-step patches, then alternating causal temporal attention and full variate attention. The Hub card lists 20 layers, model dim 1280, 16 heads, and forecast patches of 64. - Contiguous patch masking fills the horizon in one pass. The model emits 9 quantiles (10th through 90th percentile) per target step. - Google Research reports TimesFM-3 as the top-ranked pretrained foundation model on Gift-Eval, FEV-Bench, and TIME for point and probabilistic metrics, in a comparison set that includes Chronos-2, Toto 2.0, and TimesFM-2.5. Univariate mode is already competitive on those plots; multivariate mode is the further step. - Source in the timesfm repo is Apache-2.0, as are weights through 2.5. TimesFM 3.0 default weights use TimesFM Non-Commercial License v1.0: research, testing, and evaluation, not commercial or production use.

How to read this: "state-of-the-art" and the three-benchmark ranks are Google Research claims on public leaderboards. The 3.0 Hub dump is a research checkpoint with a tighter license than prior TimesFM weights.

Frontier Flux is independent coverage. We did not train or host TimesFM-3.