TimesFM-3 weights land on Hugging Face

Google on August 31 put TimesFM-3 on Hugging Face. It is a 330-million-parameter forecasting model that reads many related time series at once, and in Google’s own benchmark charts it beats Amazon’s Chronos-2, Salesforce’s Moirai 2.0 and Datadog’s Toto 2.0. The pretrained weights ship under a non-commercial license that rules out production use for now, so a forecasting pipeline that serves a live product cannot run them, while TimesFM-2.5, Chronos-2 and Toto 2.0 all remain under Apache 2.0, The New Stack reports.

The restriction covers the weights, not the code. The TimesFM-3 source code stays under Apache, but Google says that “for the time being, TimesFM 3.0 pretrained weights are distributed under the separate timesfm-non-commercial-license-v1.0 license and are restricted to non-commercial, non-production use. Commercial or production use of the default pretrained weights is not permitted.” Google has not said when “the time being” ends.

Google is also putting these models into its own products. It plans to replace TimesFM-2.5 as the model behind BitQuery’s AI.FORECAST command. The New Stack describes non-commercial weights as becoming “a bit of a trend” among model builders. It reads the combination of locked-down state-of-the-art weights and a paid route through Google’s data warehouse as “a pretty clear signal of where these labs think the money is in the long run.”

What Google measured

The benchmark results come from Google. It ran TimesFM-3 on three suites: Salesforce’s Gift-Eval, Amazon/AutoGluon’s FEV-Bench and Time. The New Stack says the new model came out ahead of the others “often by a significant margin.” Google’s own previous model sits at the bottom of the same charts. TimesFM-2.5 was state of the art when it launched in September 2025 and has already fallen to last place less than a year later.

Many series at once

Earlier TimesFM releases forecast one series at a time. According to the project’s PyPI page, the first open checkpoint, timesfm-1.0-200m, was univariate and produced only point forecasts. The 500m checkpoint released on December 30, 2024 was univariate too. It added fine-tuning and what the page calls ”~zero-shot covariate support with external regressors”, which meant outside signals were fed in alongside a single series. TimesFM-3 is Google’s first model pre-trained natively on multiple time series, with zero-shot generalization, so it needs no training on the user’s own data. It can forecast several related series in parallel and draw on historical data such as past foot traffic.

Google research scientists Ayush Jain and Rajat Sen give the reasoning in the announcement: “most real-world forecasting problems are inherently multivariate.” Their example is a retailer’s sales forecast. They say it should also use “sales of related products (e.g., ice cream cones, syrups), historical foot traffic, and known future events like weather forecasts, promotions, and holidays.”

TimesFM-3 is still a decoder-only transformer. It cuts each series into patches of 32 data points and treats each patch roughly the way a language model treats a token. What is new is that two kinds of attention layers alternate. One looks back along a single series and is strictly causal, so the model never sees values it should not know yet. The other looks sideways across all series at the same moment. That is how a promotion on one product line can shift the forecast for another.

Decoding has changed too. Previous versions generated the forecast one patch at a time, which Google’s researchers say added latency and let errors compound. TimesFM-3 appends masked placeholder tokens for the whole forecast horizon and fills them all in one forward pass. It does this with 330 million parameters, fewer than the 500 million of the TimesFM-2.0 checkpoint.