Google TimesFM-3 tops forecast benchmarks, non-commercial
TL;DR
- TimesFM-3 ranks first on Gift-Eval, FEV-Bench, and Time among pre-trained foundation models, but TSFM.ai flags that average-rank metrics obscure per-task effect sizes and results require independent reproduction on domain data.
- The licensing break from prior versions is concrete: TimesFM-2.5 shipped Apache 2.0; TimesFM-3 weights carry a new timesfm-non-commercial-license-v1.0 that explicitly prohibits production deployment and third-party mirroring.
- Non-autoregressive decoding generates the full forecast horizon in a single forward pass, a structural change from earlier TimesFM generations that decoded the horizon sequentially.
Google Research on Monday released TimesFM-3, a 330-million-parameter decoder-only foundation model that forecasts multiple time series jointly and takes the top average rank on three major zero-shot benchmarks. It ships under a non-commercial license that keeps it out of production systems.
The pitch is that forecasting in the wild is rarely one series in isolation. "Most real-world forecasting problems are inherently multivariate: where multiple time series and auxiliary external features jointly impact the future forecast of a time series," authors Ayush Jain and Rajat Sen write. Earlier versions could not do it. "Up until TimesFM-2.5 (released in September 2025), our models were strictly limited to univariate forecasting." The new release adds native support for multiple targets, past covariates, and past-future dynamic covariates such as promotion schedules or weather forecasts.
The architectural shift swaps the iterative decoder for what the team calls "Contiguous Patch Masking to generate the entire forecasting horizon in a single forward pass." Patches of 32 time steps run through alternating causal temporal attention (within-series) and full variate attention (cross-series), producing 9 quantiles per step for probabilistic forecasts.
The headline claim, per the blog: "On all three benchmarks, TimesFM-3 is the top-ranked model in terms of both point and probabilistic forecasting metrics among all pre-trained foundation models." MarkTechPost breaks that down as rank #1 on fev-bench across 100 real-world tasks and rank #1 on TIME across 50 domain datasets and 98 evaluation tasks, ahead of Chronos-2 and the Toto 2.0 family.
The catch is the license. As The New Stack headlined it, "You can't use it at work (yet)." The Hugging Face weights "are restricted to non-commercial, non-production use," while the code repository stays Apache-2.0. That puts TimesFM-3 in a different bucket from the permissively-licensed drops we've been tracking on the open source beat (255 alerts in the last 90 days): fair game for benchmarking today, not for shipping behind a forecast API. BigQuery integration is promised "in the coming weeks."
What others are reporting
-
The New Stack Read →
Dev and engineering audience angle on the deployment gap: top benchmark scores do not translate to workplace use until the BigQuery commercial route ships.
-
TSFM.ai Read →
Specialist time-series outlet applying a critical evaluation framework: average rank obscures effect size, covariate gains depend on dataset structure, and independent reproduction is needed.
Google's new 330-million-parameter model is pretrained for joint multivariate forecasting, accepts both historical-only and future-known covariates.
-
DataNorth AI Read →
Practitioner focus: benchmarks need validation on company-specific data, and the BigQuery integration matters more to businesses than direct weight access.
TimesFM-3 can forecast several related target series together and use additional variables that may improve the prediction.
-
explainx.ai Read →
Leads with the license break as the main story: non-commercial restriction blocks self-hosting and third-party mirroring in a way prior Apache 2.0 terms did not.
Accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models.
Originally reported by research.google
Read the original article →Original headline: Google Research Ships TimesFM-3, a 330M-Param Zero-Shot Foundation Model for Multivariate Forecasting