Google's TimesFM-3 Tops Forecasting Benchmarks, but Commercial Use Blocked for Now

Google has rolled out TimesFM-3, a 330-million-parameter forecasting model trained on over a trillion time-series data points that significantly outperforms rival systems from Amazon, Salesforce, and Datadog. The model is available on Hugging Face but only for non-commercial work—a licensing restriction that signals Google's strategy for monetizing advanced AI.

Why Multivariate Forecasting Matters#

Time-series models work similarly to large language models, predicting what comes next—except for data instead of text. In real-world business scenarios, forecasts rarely depend on a single data stream. Predicting ice cream sales, for instance, requires factoring in related product lines, past foot traffic, weather forecasts, and upcoming promotions.

TimesFM-3 is Google's first natively pretrained system built to handle multiple interconnected time series simultaneously using what's called zero-shot generalization. This means it can work with new data without requiring task-specific fine-tuning.

Performance Gains Across Benchmarks#

Google's benchmarking tests show TimesFM-3 outperforming competing models—including Salesforce's Moirai 2.0, Amazon's Chronos-2, and Datadog's Toto 2.0—often by substantial margins. Remarkably, TimesFM-2.5, which held state-of-the-art status when released just months ago, now ranks at the bottom of these same benchmarks, illustrating how rapidly the field is advancing.

Technical Architecture Improvements#

Like earlier versions, TimesFM-3 uses a decoder-only transformer architecture that chunks time series into 32-point patches treated as tokens. Two key innovations distinguish it:

• Dual attention layers: One scans backward through a single series (maintaining causal integrity), while the other looks sideways across all series at the same time, allowing cross-series information to influence forecasts
• Parallel decoding: Rather than generating forecasts sequentially, which introduces latency and error accumulation, TimesFM-3 fills all forecast positions in a single forward pass

The Commercial License Barrier#