Google's TimesFM‑3 forecasts multiple time series at once
September 1, 2026
The open model jointly processes sales, weather or sensor data and produces a complete forecast in one pass. Google reports top rankings on three public benchmarks.
What this is about
Google Research released TimesFM‑3 on 31 August 2026, an openly available foundation model for time series. It does not predict the next sentence but future measurements: sales, electricity demand, machine conditions or patient data. The major change from TimesFM‑2.5 is its ability to process multiple related series together.
TimesFM‑3 has 330 million parameters and, according to Google, was pretrained on a mixture of real and synthetic data containing more than one trillion time points. The weights and model card are available on Hugging Face, while the source code is on GitHub. Companies and researchers can therefore inspect and run the model themselves, although Google's performance claims still need independent testing.
What TimesFM‑3 actually does
A conventional single-series forecast might consider only historical ice-cream sales. TimesFM‑3 can also include related products, visitor counts, planned promotions, holidays and weather forecasts. Some variables are known only for the past; others, such as a holiday calendar, are also known for the future.
Technically, the model divides time series into patches of 32 steps. Its attention alternates between the time axis and relationships among different series. Instead of generating the forecast horizon piece by piece, it calculates all future patches in one pass. For each point in time it returns nine quantiles from the 10th to the 90th percentile. This provides not just an estimate but also a range representing uncertainty.
Why it matters
Time series sit behind many everyday decisions: How much stock should be ordered? When might a machine fail? How heavily will an electricity grid be loaded tomorrow? Strong models have often required separate tuning for each task. TimesFM‑3 is intended to provide useful forecasts without task-specific training. That lowers the entry barrier for teams without a large machine-learning department.
Google reports the best average rank among pretrained models for TimesFM‑3 on Gift‑Eval, FEV‑Bench and the TIME leaderboard, for both point estimates and probabilistic forecasts. That is a meaningful signal, but it does not prove superiority in every business. Public benchmarks represent incomplete, noisy company data only imperfectly.
In plain language
A conventional forecast is like looking at umbrella sales from the past few weeks. TimesFM‑3 also looks at the weather forecast, school holidays, foot traffic and raincoat sales. It packs all those clues into one suitcase and returns not only a number but also a range showing how uncertain that number is.
A practical example
A bakery chain wants to plan croissant demand for the next 28 days. It provides daily sales from 20 branches, historical visitor numbers, known holidays, planned discounts and weather forecasts. TimesFM‑3 produces a forecast for all branches and days at the same time.
For one branch, the model might provide 420 croissants as the median for Saturday and an uncertainty range of 350 to 510. The purchasing team should not blindly order 420. When uncertainty is high, it can plan cautiously, compare the forecast with local events and continuously measure actual sales against predictions.
Scope and limits
First, the architecture, training volume and benchmark results come from Google's own publication. Independent reproductions of TimesFM‑3 were not yet available at release time.
Second, even a large model cannot identify events that appear neither in the input data nor in known future signals. A strike, broken refrigeration unit or sudden ban can make any forecast useless.
Third, open weights do not automatically mean simple or risk-free deployment. Teams must assess data quality, computing costs, licence terms, privacy and systematic errors themselves. In medicine, finance or critical infrastructure, a forecast must not replace human approval.
SEO & GEO keywords
TimesFM‑3, Google Research, time-series forecasting, multivariate forecasting, zero-shot forecasting, open-weight AI, Hugging Face, Gift‑Eval, FEV‑Bench, TIME leaderboard, demand planning
💡 In plain English
TimesFM‑3 combines multiple numerical series and known future data into one forecast. The open model can start without task-specific training, but it must be tested on local data before real decisions are made.
Key Takeaways
- →TimesFM‑3 jointly processes multiple target series and additional variables.
- →The model has 330 million parameters and Google says it was pretrained on more than one trillion time points.
- →It produces the full forecast horizon in one pass rather than step by step.
- →Google reports top rankings on three public forecasting benchmarks.
- →Weights and code are public, but independent real-world testing is still necessary.
FAQ
Is TimesFM‑3 a language model?
No. It uses a transformer architecture, but it forecasts numerical series rather than text.
Must it be trained for every task?
Google describes TimesFM‑3 as a zero-shot model. It can start without task-specific training, but it should be validated on local data.
What data can it combine?
It supports multiple targets, variables known only in the past and future signals such as holidays or planned promotions.
Is the model publicly available?
The weights and model card are on Hugging Face, and the code is on GitHub. Users must check the applicable licence terms.