Forecasting is part of daily life in business. Retailers must decide how much stock to order. Energy teams must predict tomorrow’s power demand. Finance teams must plan cash flow months ahead. For years, doing this well meant building and tuning a separate model for every dataset. TimesFM 3.0 changes that idea. Released by Google Research in August 2026, it is a pretrained forecasting model that can predict data it has never seen, with no extra training.

This guide explains what the model is, how it works inside, where you can use it, and which limits you should know about. Whether you are a beginner curious about AI or a data professional comparing tools, each section is written in plain language. If you want to build a career around models like this, a credential such as the Certified Artificial Intelligence (AI) Expert program can help you learn the foundations of modern AI systems.

What Is TimesFM 3.0?

TimesFM stands for Time Series Foundation Model. A time series is simply a list of numbers recorded over time, such as daily sales, hourly website visits, or monthly rainfall. TimesFM 3.0 is the third generation of Google Research’s family of models built to forecast these lists.

The model has 330 million parameters, which are the adjustable values a neural network learns. That is small compared with large language models, which often have billions. Google Research pretrained it on a mix of real and synthetic data containing more than 1 trillion time points. Because it has already seen a huge variety of patterns, it can forecast new series “zero-shot,” meaning without being trained on your own data first.

The most important upgrade is native multivariate forecasting. Earlier versions, up to TimesFM 2.5 released in September 2025, could only look at one series at a time. TimesFM 3.0 can read several related series together and can also use outside signals like promotions, weather, and holidays.

Quick Facts

Developer: Google Research
Release: August 2026
Size: 330 million parameters
Training data: More than 1 trillion time points, real and synthetic
Type: Decoder-only transformer for time series
Main skill: Zero-shot univariate and multivariate forecasting
Where to use it: BigQuery, GitHub, and Hugging Face

Why Time Series Forecasting Needed a New Approach

To see why this model matters, it helps to look at how forecasting used to work. Traditional tools such as ARIMA or exponential smoothing needed a specialist to pick settings for each dataset. Machine learning models such as gradient boosting or deep neural networks often needed lots of historical data, feature engineering, and retraining whenever conditions changed.

That approach is slow and costly when a company has thousands of products, stores, or sensors. Each one needs its own tuning. Small teams often skip forecasting altogether or settle for simple averages.

Foundation models offer a different deal. Large language models learned to handle many writing tasks by reading huge amounts of text. TimesFM applies the same idea to numbers. It studies a massive library of series from many fields, then carries that general knowledge to your problem. Professionals who want to build such systems themselves can deepen their skills through the Certified Artificial Intelligence (AI) Developer path, which covers practical model building and deployment.

From Single Series to Real-World Complexity

Real forecasting problems are rarely about one line of numbers. Google Research gives a simple example in its announcement. If you want to forecast ice cream sales for a retail chain, past sales alone do not tell the full story. You also want sales of related items like cones and syrups, foot traffic, weather forecasts, and planned promotions. TimesFM 3.0 was built to use exactly this kind of information.

How Does TimesFM 3.0 Work?

Under the hood, TimesFM 3.0 uses a decoder-only transformer, the same broad family of architecture behind many modern language models. According to the Google Research blog, the process can be understood in four steps.

Step 1: Patching and Normalizing the Data

The model does not read one number at a time. It groups contiguous data points into patches of 32 time steps. Each patch acts like a “word” in a sentence, which makes processing faster and helps the model see short-term shapes in the data.

Next comes normalization per series. A series that counts millions of dollars and another that counts a few units are very different in scale. Normalizing puts them on a similar footing so the model can compare patterns rather than raw size.

Step 2: Building Tokens with Covariates

A covariate is any extra series that helps explain the target you want to predict. TimesFM 3.0 handles three kinds of input:

Multiple targets: Several related series forecast together, such as different ice cream brands.
Past covariates: Signals known only for the past, such as earlier foot traffic.
Past-future covariates: Signals known ahead of time, such as holidays, weather forecasts, or planned promotions.

For target and past-covariate series, each token comes from a single patch. For past-future covariates, the model uses a lookahead design. Each token joins the current patch with upcoming patches, so the model can “see” known future events.

Step 3: Alternating Attention

After tokenization, the data passes through a stack of transformer layers arranged like a grid. Two kinds of attention alternate through the layers:

Causal temporal attention: Each token looks backward across time within its own series. The word “causal” means it cannot peek at the future of that series, which prevents data leakage.
Full variate attention: At each moment, tokens look across all the series. This is how the model learns links between series, such as how a promotion on one product changes sales of another.

By alternating these two views, the model blends time patterns with cross-series relationships. The model card on Hugging Face lists 20 transformer layers with a model dimension of 1280 and 16 attention heads.

Step 4: Forecasting in a Single Pass

Older TimesFM versions produced forecasts one patch at a time, feeding each output back in as input. That loop adds delay and lets small errors pile up. TimesFM 3.0 uses a technique called Contiguous Patch Masking to produce the entire forecast horizon at once.

The model adds blank placeholder tokens for the future next to the observed history. Target and past-covariate series stay masked in the future because their values are unknown. Past-future covariates stay visible because those values are known. The attention layers then fill in every future patch at the same time, with no loop.

The output is not just one line. The model predicts nine quantiles, from the 10th to the 90th percentile, for each target at every future step. Quantiles show a range of likely outcomes, so you get a view of uncertainty instead of one fragile guess. If you work in this field professionally, a general Tech Certification can broaden your skills across the AI, data, and cloud tools that models like this depend on.

Key Features of TimesFM 3.0

Several traits set this model apart from older forecasting tools.

Zero-shot accuracy. You can send your data to the model and receive forecasts without training anything. This lowers the barrier for teams without machine learning specialists.

Native multivariate support. Instead of bolting extra features onto a single-series model, TimesFM 3.0 was pretrained for multivariate tasks from the start.

Probabilistic forecasts. Nine quantiles give a full picture of risk, which helps in planning stock levels, staffing, or budgets.

Fast inference. Producing the whole horizon in one pass reduces delay and avoids compounding errors.

Compact size. At 330 million parameters, it is far lighter than a large language model, so storage and memory needs stay modest.

TimesFM 3.0 vs TimesFM 2.5

Feature	TimesFM 2.5	TimesFM 3.0
Forecasting scope	Univariate only	Univariate and multivariate
Target columns	Single	Single or multiple
Past-only covariates	Not supported	Supported
Past and future covariates	Not supported	Supported
Decoding	Patch by patch	Whole horizon in one pass
Parameters	About 200 million	330 million

Google Cloud confirms that existing BigQuery workflows can upgrade by choosing the new model name, so you do not need to rewrite your queries to gain the accuracy boost.

Benchmark Performance

Google Research tested TimesFM 3.0 on three public benchmarks: GIFT-Eval, FEV-Bench, and TIME. On all three, it ranks first among pretrained foundation models for both point forecasts and probabilistic forecasts. The comparison included recent models such as Chronos-2 and the Toto 2.0 family, along with TimesFM 2.5.

One detail is worth noting. Google reports two results for the new model. In univariate mode, where the model ignores covariates and other series, it already matches or beats competitors. In full multivariate mode, it improves again. That suggests the gain comes both from better base forecasting and from the new ability to use extra information.

A useful reminder for any reader: benchmark rankings come from public datasets and Google’s own reporting. Your own data may behave differently, so test the model on your own history before relying on it.

Real-World Use Cases

Because it works on any numeric sequence, TimesFM 3.0 fits many industries.

Retail: Forecast demand while accounting for promotions, holidays, and weather. Google’s blog example shows a forecast that anticipates roughly a 20% sales lift on days with planned promotions, while a univariate forecast misses it.
Transportation: The BigQuery announcement uses NYC Citi Bike data to forecast commuter and casual rides together, capturing shared weather sensitivity and different weekend patterns.
Energy and utilities: Predict load using temperature forecasts and calendar effects.
Finance and operations: Plan cash flow, traffic, or support ticket volume.
Monitoring: Spot unusual behavior by comparing actual values against predicted ranges. BigQuery also supports anomaly detection with this model.
Marketing: Estimate campaign impact and seasonal demand. Teams planning data-driven campaigns can pair such tools with a Marketing Certification to connect forecasts with strategy.

How to Use TimesFM 3.0

There are two main routes, depending on your goals.

Option 1: Use It in BigQuery

The easiest path is Google Cloud BigQuery. Google announced a preview of TimesFM 3.0 there on October 12, 2026. You call it through the AI.FORECAST function in SQL and set the model to ‘TimesFM 3.0’. No model hosting or training is needed.

For multivariate work, the function accepts a few key inputs:

TARGET_COLS: The series you want to predict
PAST_COVARIATE_COLS: Signals known only for history
FUTURE_COVARIATE_COLS: Signals known for the future too

The data goes in one table. Historical rows are fully filled in. Future rows contain only the known future covariates, with target and past-only columns left as NULL. The result returns a forecast value plus lower and upper prediction bounds for each target. You can also ask for forecasts in plain English through BigQuery Conversational Analytics, which writes the query for you.

Option 2: Run the Open Checkpoint

Developers can download the checkpoint named google/timesfm-3.0-pytorch from Hugging Face and use the code in the official GitHub repository. This suits research, experiments, and learning. The repository includes a forecaster class that supports univariate or multivariate input with past-only and past-future covariates. Before choosing this route, read the licensing section below.

Licensing: What You Can and Cannot Do

Licensing is an important detail. According to the official GitHub repository, the TimesFM 3.0 pretrained weights come under a separate non-commercial license and are restricted to non-commercial, non-production use. The same page states that commercial and production use is fully permitted through authorized Google Cloud services, including BigQuery ML, under the Google Cloud Terms of Service.

In simple terms, students and researchers can experiment with the downloaded weights. Companies that want production forecasts should use the managed Google Cloud route. Always check the current license text before you build a product, because terms can change.

Limitations to Keep in Mind

No forecasting model is perfect, and honest planning means knowing the weak spots.

Forecasts are not guarantees

The future can shift because of events no model has seen. The prediction intervals exist to remind you that uncertainty is normal.

Covariates need to be reliable

If you feed in a poor weather forecast or an inaccurate promotion plan, the forecast will reflect that.

Zero-shot is a starting point

For specialized or unusual data, results may still improve with domain knowledge, careful data cleaning, or comparison against simpler baselines.

The BigQuery feature is in preview

Preview features can change, so check the documentation before building critical systems.

Test on your own data

Public benchmark leadership does not promise the best result on every private dataset.

The Future of Time Series Foundation Models

TimesFM 3.0 shows where forecasting is heading. The field is moving from many small custom models toward one general model that handles many tasks well. The jump from single-series input to native multivariate input also brings forecasting closer to how real businesses think, where many factors act together.

For learners, this is a good time to build skills in AI, data, and cloud platforms. Courses, hands-on practice with public datasets, and certifications can all help you move from curiosity to confidence. A Marketing Certification can also help non-technical professionals understand how forecasting supports planning and growth.

Conclusion

TimesFM 3.0 is Google Research’s multivariate time series foundation model. It has 330 million parameters, learns from more than a trillion time points, and forecasts new data without retraining. Its patching, alternating attention, and single-pass decoding allow it to read several series and outside signals at once and return a full range of likely outcomes. It leads three major public benchmarks, and you can try it today in BigQuery or through the open checkpoint for non-commercial work. Start with a small dataset, compare the results with a simple baseline, and judge it by your own numbers.

Official Sources

FAQs

1. What is Google TimesFM 3.0?

Google TimesFM 3.0 is a time-series foundation model developed by Google Research to predict future values using historical data and related information. It supports multivariate forecasting, allowing businesses to analyse multiple interconnected data points and generate predictions without training a separate model for every forecasting task.

2. What does TimesFM stand for?

TimesFM stands for Time Series Foundation Model. It is a pretrained artificial intelligence model designed to understand patterns in time-dependent data and forecast future trends across different industries.

3. How does Google TimesFM 3.0 work?

TimesFM 3.0 analyses historical time-series data, identifies patterns, and uses a Transformer-based architecture to generate future predictions. It can also consider related variables, such as weather, promotions, holidays, and other external factors, to improve forecasting results.

4. What makes TimesFM 3.0 different from previous versions?

TimesFM 3.0 introduces native multivariate forecasting, allowing it to predict multiple related time series simultaneously. Earlier versions primarily focused on univariate forecasting, where predictions were based on the historical values of a single series.

5. What are the key features of Google TimesFM 3.0?

Major features include:

Multivariate and univariate forecasting.
Support for multiple related target variables.
Historical and future covariate support.
Zero-shot forecasting capabilities.
Probabilistic predictions.
Transformer-based architecture.
Integration with Google BigQuery.

6. What is multivariate time-series forecasting?

Multivariate time-series forecasting predicts future values by analysing multiple related variables together. For example, a retailer can forecast product sales using historical sales, customer traffic, promotional activities, and seasonal trends instead of examining sales data independently.

7. What is zero-shot forecasting in TimesFM 3.0?

Zero-shot forecasting allows TimesFM 3.0 to generate predictions for new datasets without requiring task-specific model training or fine-tuning. Because the model has already learned patterns from large-scale time-series data, it can apply its pretrained knowledge to different forecasting problems.

8. How many parameters does Google TimesFM 3.0 have?

TimesFM 3.0 contains approximately 330 million parameters. Google Research reports that it was pretrained on a combination of real-world and synthetic time-series data containing more than one trillion time points.

9. What is the role of Transformer architecture in TimesFM 3.0?

The Transformer architecture helps TimesFM 3.0 identify relationships and patterns across historical data. Its attention mechanisms analyse temporal relationships within individual series and correlations between different variables, helping the model generate more informed predictions.

10. What are covariates in TimesFM 3.0?

Covariates are additional variables that can influence the target being forecast. TimesFM 3.0 supports two main categories:

Past covariates: Historical information that is not available for future dates, such as previously recorded customer traffic.
Future covariates: Information known in advance, such as scheduled promotions, holidays, and calendar events.

11. How does TimesFM 3.0 help businesses predict future sales?

Businesses can use TimesFM 3.0 to analyse historical sales alongside related factors such as seasonal demand, promotions, and customer activity. This can help retailers estimate future product demand, manage inventory, and plan marketing campaigns more effectively.

12. Can Google TimesFM 3.0 forecast multiple variables simultaneously?

Yes. TimesFM 3.0 supports multiple target variables in a single forecasting process. For example, a business can forecast sales for different product categories simultaneously while considering shared influences such as holidays and promotional campaigns.

13. What industries can benefit from TimesFM 3.0?

TimesFM 3.0 can support forecasting applications across several industries, including:

Retail and e-commerce.
Banking and financial services.
Healthcare.
Manufacturing.
Energy and utilities.
Supply chain management.
Transportation and logistics.

14. How is Google TimesFM 3.0 integrated with BigQuery?

Google integrates TimesFM 3.0 into BigQuery through the AI.FORECAST SQL function. Users can generate forecasts directly from datasets without building and managing a separate forecasting model. The integration also supports anomaly detection and forecasting evaluation workflows.

15. Can users access TimesFM 3.0 without advanced machine learning knowledge?

Yes. Through Google BigQuery, users can run forecasting tasks using SQL rather than developing an entire machine learning pipeline. BigQuery Conversational Analytics also allows users to request forecasts using natural-language instructions.

16. What are probabilistic forecasts in TimesFM 3.0?

Probabilistic forecasts provide a range of possible future outcomes instead of presenting only one predicted value. TimesFM 3.0 generates multiple quantile predictions, helping users understand forecast uncertainty and evaluate different potential scenarios.

17. How accurate is Google TimesFM 3.0?

Google Research reports that TimesFM 3.0 achieved leading performance across three major public forecasting benchmarks: GIFT-Eval, FEV-Bench, and TIME. However, forecasting accuracy varies depending on the dataset, industry, prediction horizon, and quality of the input information.

18. Is Google TimesFM 3.0 open source?

Google has published TimesFM 3.0's source code and pretrained model weights through its GitHub and Hugging Face repositories. However, there is an important licensing distinction: the TimesFM 3.0 pretrained weights are distributed under a non-commercial licence, while commercial and production use is permitted through authorised Google Cloud services such as BigQuery ML.

19. What are the limitations of TimesFM 3.0?

Although TimesFM 3.0 offers advanced forecasting capabilities, it cannot guarantee accurate predictions in every situation. Unexpected market changes, incomplete datasets, unusual events, and poor-quality input variables can affect results. Businesses should evaluate forecasts against actual outcomes before using them for important decisions.

20. What is the future of AI-powered forecasting with TimesFM 3.0?

TimesFM 3.0 represents an important development in AI-powered predictive analytics. Its ability to analyse multiple interconnected variables could help organisations improve demand forecasting, financial planning, operational efficiency, and resource management. Integration with cloud analytics platforms may also make advanced forecasting more accessible to businesses.