Google's TimesFM-3 Tops Forecasting Benchmarks Under a Non-Commercial License

Google released TimesFM-3 over the weekend, the newest version of its time-series foundation model, and the benchmarks say it is the best zero-shot forecaster you can download. The catch is in the license: as MarkTechPost’s coverage of the Google Research announcement notes, the weights are restricted to non-commercial, non-production use, a significant step back from TimesFM-2.5’s Apache 2.0 terms.

What Actually Changed in the Model

TimesFM-3 is a 330M-parameter model, small by frontier standards, and its headline capability is native multivariate forecasting. TimesFM-2.5, released a year ago, handled only univariate series; version 3 forecasts many related signals jointly and accepts both past-only and future-known covariates, which is how real forecasting problems actually look: demand per SKU alongside price, weather, and calendar effects, predicted together rather than one series at a time. The weights are on Hugging Face and the code is on GitHub, so evaluation is a five-minute job.

The License Is the Real Signal

The non-commercial restriction deserves more attention than the benchmark win. TimesFM-2.5 shipped under Apache 2.0 and was widely embedded in production forecasting stacks; version 3 is benchmark-it-only. Google Cloud’s own BigQuery ML TimesFM integration sits in preview, which makes the strategy legible: give researchers and enterprises a reason to evaluate Google’s model, then funnel production usage toward the managed offering where the commercial license lives. Open weights as a funnel, not a gift. Teams that built on 2.5’s generous terms should read this as a reminder that open-weight licenses are revocable per release, not permanent commitments, and plan exit paths accordingly.

The Broader Pattern

This lands in the same month Meta shipped a 30B open-weights model under Apache 2.0, a reminder that the open-weight landscape is splitting into genuinely open releases and openly inspectable ones. For ML engineers doing forecasting work, the practical play is straightforward: benchmark TimesFM-3 now, keep 2.5 in production until the licensing picture changes, and watch whether Google reopens the terms the way it eventually did for other restricted releases.