On August 31, 2026, Google Research released TimesFM-3, a 330-million-parameter time-series foundation model that can forecast several related series in one forward pass. That is the first TimesFM release trained natively for multivariate forecasting. If you have been searching for predictive intelligence that actually uses promotions, weather, and sibling SKUs instead of a single line chart, this is the news that matters.
This post is for one search: predictive intelligence. The only product link is Predictive Data Intelligence. If your demand, churn, or capacity forecasts still ignore related series and known future events, that page is where we scope a production system on your data.
What Google actually shipped on August 31
Keep the press summary short. The research post is specific.
- TimesFM-3 has 330 million parameters and was trained on more than one trillion time points.
- It jointly predicts multiple coevolving targets, with point and quantile forecasts for each.
- It accepts past covariates (history only) and past–future covariates (known future signals such as promotions or weather).
- It generates the full horizon in a single pass via contiguous patch masking, instead of patch-by-patch autoregression.
- Google reports best average rank among pretrained foundation models on Gift-Eval, FEV-Bench, and TIME for both point and probabilistic metrics.
Earlier TimesFM versions stopped at univariate history. TimesFM-2.5 (September 2025) still could not natively use related series or known future events. TimesFM-3 closes that gap in the foundation-model layer.
The license catch most demos skip
Public weights are on GitHub and Hugging Face. Google currently limits those weights to non-commercial and non-production use. BigQuery integration is coming in the following weeks; until then, Teams can experiment, not ship.
That is the line between a research release and predictive data intelligence you can put on a planning calendar. A leaderboard win on Gift-Eval does not replace a backtest on your own SKUs, stores, or sensors. If the public checkpoint cannot go into production, you still need a model stack, monitoring, and a retraining path that your counsel will accept.
Why multivariate is the real product decision
Google’s ice-cream example is useful because it is boring. Past sales alone miss the story. Related products, foot traffic, weather, and a promotion calendar change the forecast. A univariate model draws a weekly pattern forward. A multivariate model that sees the promotion schedule can anticipate a sales bump on those days.
Most business forecasting problems look like that example:
- Retail demand with promotions, weather, and sibling SKUs
- Churn risk with usage, tickets, and billing history
- Manufacturing failure risk with vibration, temperature, and load
- Staffing and capacity with booked events and seasonal traffic
If your current “AI forecast” only reads one column, TimesFM-3 is evidence that the research community has moved on. Your ops team should move with it, with data you control.
What predictive data intelligence actually includes
On Predictive Data Intelligence we treat this as an end-to-end build, not a notebook drop. A useful scope usually covers:
- Data collection and integration from CRM, ERP, web analytics, and external APIs into one warehouse shape.
- Feature engineering that keeps promotions, holidays, and related series usable as covariates.
- Model training and validation with backtests on your horizon, not only public benchmarks.
- Deployment and continuous learning with monitoring, drift checks, and a retraining path.
Foundation models like TimesFM-3 can speed evaluation. They do not remove the need for business metrics: forecast error on the decisions you make, calibration of uncertainty, and downstream inventory or staffing impact.
How to use TimesFM-3 without fooling yourself
If you want a practical week-one plan after the August 31 release:
- Pull 12–24 months of history for the series you actually plan against.
- Add at least one known-future covariate you already schedule (promotions, holidays, maintenance windows).
- Compare univariate vs multivariate error on a held-out month that includes a real promotion or shock.
- Measure the business outcome (stockouts, excess inventory, missed shifts), not only MASE.
- Decide early whether you will stay on a managed forecast API, fine-tune under a commercial license when available, or ship a custom stack on your data.
Google itself says even univariate TimesFM-3 is competitive, and multivariate mode improves further when cross-series information and covariates exist. Your test should show whether those gains appear on your series.
When to stop reading papers and start a build
You are ready for a predictive build when three things are true:
- You have a named decision that repeats (buy, staff, retain, maintain).
- You have enough history to backtest that decision horizon.
- Someone owns the action that follows a prediction (buyer, CSM, maintenance lead).
If you only want a playground score on Gift-Eval, wait for the BigQuery path and keep reading. If a wrong forecast already costs money every week, scope the system on Predictive Data Intelligence and treat TimesFM-3 as one candidate in the bake-off, not the whole product.
Bottom line
TimesFM-3, released August 31, 2026, is a real step: multivariate, zero-shot, top-ranked on the public foundation-model benches Google published. The public weights are not a production license. The search that should bring you here is predictive intelligence, and the page that owns the next step is Predictive Data Intelligence.