Google released TimesFM-3, a 330-million-parameter time-series forecasting model trained on more than a trillion real-world and synthetic data time points, the company announced Monday. The model is now accessible on Hugging Face under a non-commercial license. This represents Google's first forecasting model designed from the ground up to process multiple time series simultaneously with zero-shot generalization.
In benchmarks shared by Google, TimesFM-3 surpassed all competing models, often by substantial margins, according to the announcement. The team evaluated performance using Salesforce's Gift-Eval, Amazon/AutoGluon's FEV-Bench, and Time benchmarks. TimesFM-2.5, which reached state-of-the-art status when it debuted in September 2025, now sits at the bottom of these same benchmarks. The model architecture relies on a decoder-only transformer that divides each time series into patches containing 32 data points, processing them similarly to how language models handle tokens.
Research scientists Ayush Jain and Rajat Sen explained in the announcement that "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." They note that past sales alone don't tell the complete story, writing that "a good forecast should also draw on sales of related products (e.g., ice cream cones, syrups), historical foot traffic, and known future events like weather forecasts, promotions, and holidays." The model can predict multiple related time series in parallel while incorporating historical information such as past foot traffic data.
What sets TimesFM-3 apart is its use of two alternating attention layers, according to Google's researchers. The first layer examines patterns backward through time within a single series, maintaining strict causality so the model can't access values it shouldn't yet know. The second layer examines patterns across all series at any given moment, enabling a promotion in one product line to influence forecasts for another. The decoding process also changed from earlier versions: rather than generating forecasts one patch at a time—which added latency and compounded errors—TimesFM-3 adds masked placeholder tokens for the entire forecast horizon and fills them all in with a single forward pass. This architectural shift addresses how quickly the field is advancing, as demonstrated by TimesFM-2.5's rapid fall from the top of performance rankings.
Google chose to release the model under a non-commercial license, marking a shift from TimesFM-2.5's Apache 2.0 license. While the TimesFM-3 source code remains under the Apache license, the pretrained weights carry restrictions: they're limited to non-commercial, non-production use, with commercial or production deployment of the default pretrained weights explicitly prohibited. Google will soon replace TimesFM-2.5 as the engine powering BitQuery's AI.FORECAST command, signaling active monetization of these models. The company is restricting access to state-of-the-art weights while simultaneously offering a paid pathway through its data warehouse—a clear indication of where forecasting model developers see long-term revenue potential. This licensing strategy suggests organizations will need to weigh the appeal of cutting-edge performance against the constraints of proprietary access terms, especially as multivariate forecasting becomes central to operational planning.