TimesFM-3 weights land on Hugging Face

Google has launched TimesFM-3, a 330-million-parameter time-series forecasting model trained on more than one trillion real-world and synthetic data points. According to The New Stack, the model is available on Hugging Face under a non-commercial license, with its default weights limited to non-commercial, non-production use.

At a glance

• TimesFM-3 was built for multivariate forecasting, combining several related series and external signals such as historical foot traffic with the series being forecast in parallel.
• Google's benchmarks place TimesFM-3 ahead of Chronos-2, Moirai 2.0, Toto 2.0 and TimesFM-2.5 on Gift-Eval, FEV-Bench and Time, with often significant margins in the reported comparisons.
• The source code remains Apache-licensed, but Google's pretrained TimesFM-3 weights cannot be used commercially or in production under the default license for the time being.

Forecasting models apply the basic logic of next-token prediction to numerical data, where the target is a future value rather than a word. The move to multivariate inputs reflects how businesses typically operate: sales, related products, foot traffic, weather, promotions and holidays can interact. TimesFM-3 therefore arrives with a stronger technical result than its predecessor, but its license limits immediate use in commercial systems.

TimesFM-3 was trained on more than one trillion time points

Time-series forecasting models estimate what comes next in sequences such as sales, demand or traffic. Earlier systems often focused on one series at a time, while newer models can ingest multiple series and auxiliary features. Google says TimesFM-3 is its first model natively pre-trained for multiple time series with zero-shot generalization, allowing related series to be forecast together while incorporating historical signals such as foot traffic.

The model was evaluated on Salesforce's Gift-Eval, Amazon and AutoGluon's FEV-Bench, and Time. In the benchmarks Google shared, TimesFM-3 outperformed Chronos-2 from Amazon, Moirai 2.0 from Salesforce, Toto 2.0 from Datadog and TimesFM-2.5. The result also illustrates the speed of development in the category: TimesFM-2.5 was state of the art when it launched in September 2025, but now sits at the bottom of the reported comparisons.

TimesFM-3 combines two attention patterns across 32-point patches

Like earlier TimesFM versions, the new model uses a decoder-only transformer. It divides each time series into patches containing 32 data points and treats those patches roughly as a language model treats tokens. TimesFM-3 then sends the tokens through alternating attention layers, adding a second dimension to the model's view of the data.

One attention pattern looks backward through time within an individual series and remains causal, preventing the model from using values that would not yet be known. The other looks across all series at the same moment, so movement in one product line can inform another. Google gives related products, promotions and other linked signals as examples of the information this design can connect.

The decoding procedure also changes. Earlier versions generated forecasts one patch at a time, which Google says increased latency and allowed errors to compound. TimesFM-3 instead appends masked placeholder tokens for the complete forecast horizon and fills them in during a single forward pass, reducing the sequential steps required to produce a forecast.

Google keeps the TimesFM-3 weights outside commercial production

The TimesFM-3 source code is distributed under the Apache license, but the pretrained weights use Google's separate timesfm-non-commercial-license-v1.0. Google states that commercial or production use of the default pretrained weights is not permitted. The weights are available on Hugging Face, giving researchers and other non-commercial users access to the released model.

TimesFM-2.5 used the Apache 2.0 license, as do Toto 2.0 and Chronos-2. Google also plans to replace TimesFM-2.5 with TimesFM-3 in BitQuery's AI.FORECAST command. That arrangement leaves the latest weights restricted while placing the model inside a service that Google can monetize, a structure already used by other forecasting-model developers.

BitQuery will get the upgradeGoogle says TimesFM-3 will soon replace TimesFM-2.5 in BitQuery's AI.FORECAST command, but the source gives no date for that change. It also does not announce a commercial license for the default weights. Until those terms change, the strongest publicly released version remains available for non-commercial and non-production work, while commercial deployments must rely on another model or a separately offered service.