Google's forecasting model wins all three benchmarks in its field — and is the first in its line you may not use at work

Google Research announced TimesFM 3.0 in its project repository on 28 August 2026, and the headline is a clean sweep: rank one overall on fev-bench across 100 real-world forecasting tasks, rank one overall on the TIME benchmark across 50 datasets and 98 tasks, and rank one among foundation models on GIFT-Eval. The third generation of the time-series model is also the first that is natively multivariate — it reads several channels at once and accepts covariates known only for the past as well as covariates known for the future. The quieter change is in the licence, and it reverses eleven years of practice in this particular line. TimesFM 1.0, 2.0 and 2.5 all published their weights under Apache 2.0. The 3.0 weights do not. They carry a separate document, the TimesFM Non-Commercial Licence v1.0, which permits testing, evaluation and academic research and then names what is excluded: any revenue-generating activity, any interaction with end users or production systems, and training, fine-tuning or distilling other models for commercial use. Redistributing the weights — or a fine-tune of them — is forbidden outright. Google states the split itself, in a boxed notice in the repository: the source code stays Apache 2.0, weights up to 2.5 stay Apache 2.0, and only the 3.0 pretrained weights are restricted, "for the time being". The licence text adds one sentence that explains where commercial users are meant to go: it applies to those who download the model, and customers reaching it through Google's API may be on other terms. TimesFM is served inside BigQuery ML, Connected Sheets and the Vertex AI Model Garden. The practical shape of this is worth naming precisely. The weights are public, free and, in the thirty days to 23 September 2026, downloaded 898,000 times. What is not public is permission to put them to work. A forecasting model is an unusual place to draw that line, because forecasting is almost by definition a commercial act: a researcher may run TimesFM 3.0 on a sales series to write a paper, but may not run it on the same series to decide how much to order.