TimesFM-3 tops forecasting benchmarks. The license blocks you.

Demand forecasting is the AI use case small operators actually need. Not a chatbot — a number that tells you how many units to buy in November. So when Google Research published TimesFM-3 and claimed the top spot on every major time-series benchmark, that was the release worth reading. Then you get to the license, and the answer to "can we ship this" is no.

What actually happened

Google Research announced TimesFM-3 on August 31, 2026: a 330-million-parameter forecasting foundation model trained on more than 1 trillion time points of real and synthetic data. The headline change from TimesFM-2.5 is native multivariate forecasting — earlier versions were univariate only, one series at a time. TimesFM-3 forecasts multiple related series together, and takes covariates: past-only features, and past-future features where you already know what is coming (a promotion, a holiday, a price change).

Google reports it ranking top among pre-trained foundation models on GIFT-Eval, FEV-Bench, and the Time leaderboard, for both point and probabilistic forecasting, ahead of Chronos-2 and the Toto 2.0 family (Google Research).

Now the part that decides whether any of that reaches your business. The weights on Hugging Face are published under the TimesFM Non-Commercial License v1.0 — non-commercial use only (Hugging Face). The previous generation, TimesFM-2.5 at 200M parameters, is Apache 2.0 (Hugging Face).

Why the TimesFM-3 license matters for your business

Read the license before you read the benchmark. That is the whole lesson, and it keeps costing teams a sprint.

The failure mode is predictable: somebody sees a leaderboard, spikes it on a laptop against last year's sales data, the forecast looks great, and it lands in a roadmap. Three weeks later a real invoice depends on a model nobody is allowed to run in production. The evaluation was never the risk. The terms were.

What to do instead, concretely. Build the pipeline, not the model. The work in a forecasting system is almost never the model — it is cleaning the sales history, aligning SKUs that got renamed, encoding promotions as covariates, deciding whether you are forecasting units or revenue, and wiring the output into a purchase order. That work is model-agnostic. Do it against TimesFM-2.5 under Apache 2.0, which you can actually run commercially, and keep the model behind an interface you can swap.

Then TimesFM-3 becomes a question of whether Google relicenses or offers commercial terms, not a rebuild. Same discipline we apply to every vendor: the forecasting logic is yours, the model is a dependency, and dependencies get swapped.

One more thing worth saying plainly — a 200M-parameter model runs on hardware you already own. Forecasting is one of the few AI workloads where a small business genuinely does not need to rent a GPU or send its sales history to anybody.

Sources: Google Research, Hugging Face.