Google Research Releases TimesFM-3, a 330M-Parameter Forecasting Model That Fills In the Future in One Pass

Google Research has released TimesFM-3, a forecasting model that predicts future values across an entire time horizon at once rather than generating them one step at a time. The model has 330 million parameters and is designed to work with time series data alongside related signals, including known future events such as sales promotions or weather forecasts. The release moves Google's time-series work from a point-by-point approach to a single-pass one, a change that matters for anyone running forecasts at scale.

The core technical shift is what the model does with the forecasting window. Traditional autoregressive forecasters predict the next value, feed it back in, and repeat, so each step depends on the last. TimesFM-3 instead emits all future time points in one forward pass. Google Research says this cuts compute time and reduces compounding errors, the drift that builds up when a model's own earlier predictions become its inputs.

Conditioning on covariates is the other half of the design. The model does not just extrapolate from a historical series; it can take in related data and events that are already known to be coming, like a scheduled discount or a weather outlook. That matters for retail and supply-chain planning, where the future is not purely a continuation of the past and a promotion calendar is often available weeks in advance. A model that ignores those signals leaves accuracy on the table.

The parameter count puts TimesFM-3 in the mid-size range for a foundation model, far smaller than the large language models Google ships to consumers but substantial for a time-series forecaster. At 330 million parameters, it is large enough to absorb broad patterns across many domains, while still being practical to serve. Google Research has not tied the release to a specific commercial product in the summary of the work, so the immediate audience is researchers and engineers who build forecasting pipelines.

The practical stakes are in cost and reliability. Forecasting at scale is often limited less by accuracy than by how many series a system can process and how quickly, and a single-pass model changes that math. Compounding error is the quieter problem: in long-horizon autoregressive forecasts, small mistakes early in the window amplify later, which is why many production systems cap how far ahead they predict. Removing the feedback loop attacks both issues at once.

What remains open is how TimesFM-3 performs against established baselines on standard forecasting benchmarks, and whether the single-pass approach holds up on the long horizons where autoregressive models struggle most. Google Research has published the model details; independent replication and head-to-head comparisons will follow. For teams weighing a switch, the questions will be accuracy per series, inference cost, and how easily known future events can be wired into an existing pipeline.