TimesFM 3.0: Google's Time-Series Forecasting Model Just Got a Major Upgrade

Time-series forecasting is one of those tasks that quietly powers a lot of decisions, from inventory planning to energy load prediction to financial risk. But building accurate forecasting models from scratch is notoriously painful. What if you could just download a pretrained model that works out of the box, the same way you might use GPT for text? That is exactly the idea behind Google Research's TimesFM.

What Is TimesFM?

TimesFM (Time Series Foundation Model) is a decoder-only foundation model built specifically for time-series forecasting. It was introduced in the paper A decoder-only foundation model for time-series forecasting, presented at ICML 2024. Instead of being trained on words, it is trained on sequences of historical data points. Once pretrained, it can forecast future values for all kinds of time series — no custom training required for each new dataset.

Welcome TimesFM 3.0

Fast forward to August 2026 (yes, the project moves fast), and Google has released TimesFM 3.0. This is the biggest update yet, and it brings three major things to the table:

Benchmark Results That Back It Up

The readme claims some pretty impressive benchmark placements:

Being at the top of all three major time-series foundation model benchmarks is a strong signal that TimesFM has become the go-to generalist model in this space.

But Watch Out for the License

Here is the catch that matters if you want to use TimesFM 3.0 in a commercial product. The source code is Apache-2.0 licensed, and model weights up to version 2.5 are also Apache-2.0. However, the TimesFM 3.0 pretrained weights are distributed under a separate non-commercial license. That means if you are planning to use the default 3.0 weights in production or for commercial purposes, you are not allowed to do so. Make sure to review the repo for license updates before you build anything serious.

Getting Started with TimesFM 3.0

Installation is straightforward if you are already comfortable with Python and PyTorch:

Or, if you want to pull the code and install locally:

Simple Univariate Forecast Example

Here is a minimal example from the readme that shows how to forecast a couple of different-length series with quantiles:

Notice that TimesFM can take two series of different context lengths in the same batch — 100 steps for one, 72 for the other — and predict the next 12 steps for both. Plus it returns nine quantiles (0.1 to 0.9), which is handy for understanding forecast uncertainty.

Multivariate Forecasting with Covariates

If you work with multiple related signals — say three different store sales channels — you can forecast them simultaneously. Here is a condensed version of the readme's example:

The model outputs a joint forecast for all three variates over the next 24 steps, plus quantiles for each variate and horizon. That kind of native multivariate support makes it much easier to forecast a system of related time series rather than treating each one in isolation.

A Look Back: What Changed in 2.5

Before 3.0, the 2.5 release (September 2025) already brought several notable improvements: it shrank the model from 500M to 200M parameters, expanded the maximum context length from 2048 steps to 16k steps, and added an optional quantile head for continuous forecasts up to 1,000 steps. It also removed the frequency indicator, simplifying the input format. If you are using the 2.5 checkpoint and want to fine-tune it, the repository now includes examples for parameter-efficient fine-tuning with HuggingFace Transformers + PEFT (LoRA), plus unit tests and an agent skill.

Where Can You Use TimesFM?

Google has also embedded TimesFM into a few first-party products:

For everyone else, the open-source version can be installed from PyPI and run with your own infrastructure.

Final Thoughts

TimesFM 3.0 looks like a genuinely big step forward for open time-series forecasting. The ability to handle multivariate series with covariates, combined with top benchmark results, makes it a powerful foundation for a wide range of forecasting problems. Just be careful with the non-commercial weight license if you plan to use it in a product.

If you are curious, head over to the TimesFM GitHub repository to try the examples yourself. It only takes a few lines of Python to start forecasting.