Google Research has introduced TimesFM-3, a 330-million-parameter foundation model for multivariate time-series forecasting, and says it ranks first on three public benchmarks — Gift-Eval, FEV-Bench and Time — on both point and probabilistic metrics. The post ends by noting the model is "now available on GitHub and Hugging Face", with BigQuery integration "landing in the coming weeks". What the post does not contain, anywhere, is the word licence.
The term that is missing from the announcement
The Hugging Face model card for google/timesfm-3.0-pytorch states the terms plainly: TimesFM Non-Commercial License v1.0, non-commercial and non-production use. The previous release, timesfm-2.5-200m-pytorch, is Apache-2.0. This is the first licence change in the series, and it is discoverable only by opening the model card.
What the common framing gets wrong
TimesFM-3 is being described as an open-source release from Google. It is not. The repository code remains Apache-2.0; the weights are not. Those are two separate documents, and conflating them turns a research preview into a production tool that does not exist. By the OSI definition this is not open source, and it is not open-weight in the sense practitioners use the term either. The practical consequence: an engineer at a forecasting-heavy business — energy, retail, finance — cannot put the current best-ranked model into a pipeline.
What is technically new
The model is natively pre-trained for multivariate forecasting rather than stitched together from univariate runs. It is a decoder-only transformer with alternating attention layers that generates the whole horizon in a single forward pass rather than autoregressively, and it emits both point forecasts and nine quantiles, the 10th through 90th percentile. The pre-training corpus mixes real and synthetic series past 1 trillion time points.
One date to get right
The GitHub tag TimesFM-3.0 landed on 28 August; the public announcement came on 31 August. The weights sat in the repository before anyone was told about them, which is worth stating explicitly rather than dating the release to the blog post. The direction of travel is the story: a line that shipped three consecutive versions under Apache-2.0 closed its weights the version it started topping leaderboards.