Google has released TimesFM-3, a forecasting model designed to produce predictions from user data without requiring fine-tuning on that data. According to the company, users can upload a CSV file and obtain forecasts directly.
Capabilities of the model
TimesFM-3 was trained on over a trillion time points. Google reports that the model ranks first on three major benchmarks — Gift-Eval, FEV-Bench, and Time — without any fine-tuning. The model can handle multiple variables simultaneously, enabling forecasts that incorporate sales, foot traffic, promotions and other series in a single output.
In a demonstration, Google used an ice-cream vendor example: the model read the promotion calendar and increased the forecast by about 20% on each promotion day. Google emphasized that the higher predictions followed from reading the schedule rather than guessing.
Technical details and availability
TimesFM-3 has approximately 330 million parameters, which Google says allows the model to run on a laptop. The project is available on GitHub, facilitating access and testing for users.
Practical implications
Google’s presentation suggests that zero-shot forecasting can reduce or remove the weeks-long model tuning that data scientists traditionally perform. With a CSV upload, organisations can receive operational forecasts without bespoke fine-tuning, which is relevant for retailers, banks and supply chains that rely on accurate time-series predictions.
Google also noted that in many forecasting projects the largest cost has not been the software itself but personnel costs — the salaries of data scientists and modelers. By enabling zero-shot forecasts, TimesFM-3 could change how forecasting work is done in practice, according to the company's framing.
Summary
TimesFM-3 aims to simplify multivariate time-series forecasting by generating predictions from CSV inputs without fine-tuning. Trained on over a trillion time points, with about 330 million parameters and top benchmark rankings, Google positions the model as a step toward faster, lower-overhead forecasting workflows.
