TimesFM is a pretrained foundation model for time-series forecasting, released by Google Research as an open Python package with downloadable checkpoints. It is aimed at data scientists, analysts and ML engineers who forecast things like demand, traffic, load or revenue across many series at once, and who would rather call one general model than fit, tune and maintain a separate model for every dataset they own.
What it does
The core promise is zero-shot forecasting: you hand the model the history of a series and it returns the continuation, without any training step of your own. The same checkpoint is meant to work across domains and sampling frequencies, which is what makes it interesting for shops that have hundreds or thousands of series and no appetite for hundreds or thousands of models.
The latest release, TimesFM 3.0, announced in the repository in August 2026, adds:
- native multivariate forecasting, so several related series can be forecast together in a single call rather than one at a time;
- flexible covariate support, both past-only covariates and covariates whose future values you already know;
- stronger zero-shot generalist behaviour, with the project claiming top performance across the major time-series benchmarks.
Output is not just a single line into the future. As the video demo shows, a call returns the point forecast together with a set of quantile levels — nine confidence levels alongside the central estimate — so you get an uncertainty band without any extra machinery.
The model is also not a lab curiosity. Google runs it inside its own products: BigQuery ML exposes it through SQL for large-scale, reliable forecasting jobs, Connected Sheets brings the same forecasts into a spreadsheet, and Vertex Model Garden offers a containerised endpoint suitable for agent-style calling.
How it works
TimesFM is a decoder-only model, described in the ICML 2024 paper "A decoder-only foundation model for time-series forecasting" (arXiv 2310.10688). The architectural bet is the one the video states plainly: it reads a numeric series much the way a language model reads text, consuming the history as context and generating what comes next. Because the knowledge lives in the pretrained weights rather than in per-dataset fitting, generalisation to an unseen series is the normal mode of operation instead of a special case.
Everything the user supplies is input, not training data. That keeps the operational story simple — load a checkpoint, pass arrays in, get forecasts and quantiles out — and it is why the model slots so easily behind a SQL function or an HTTP endpoint.
Getting started
The current weights are published on Hugging Face as google/timesfm-3.0-pytorch; earlier releases up to 2.5 live in the project's Hugging Face collection. In the repository, code for the 2.5 line sits under src/timesfm, while the 1.0 and 2.0 generations are archived in the v1 subdirectory — if you need to load those older checkpoints, the README points you at pip install timesfm==1.3.0 to get the matching package version.
The repository is Apache-2.0 licensed, Python throughout, and has been under active development since April 2024, with commits as recent as September 2026 and roughly 30,900 stars and 2,900 forks. One caveat the README states directly: this open version is not an officially supported Google product. Treat it as research code you are adopting on your own terms.
When to use it / when not
It fits well when you have many series and little labelled history per series, when new series appear constantly and cold-start forecasts matter, when you want a credible baseline in an afternoon rather than a modelling project, or when you need forecast intervals and related series handled in one call.
It fits less well when you need a vendor-supported component with a support contract — in that case the BigQuery ML or Vertex Model Garden routes are the sanctioned path, not this package. It is also overkill when you have one long, stable, well-understood series that a carefully tuned conventional model already forecasts well; a foundation model buys you breadth, and breadth is not what that problem needs.
Alternatives
The most immediate alternatives are Google's own delivery channels for the same model: BigQuery ML if your data already lives in the warehouse and you would rather write SQL, Connected Sheets if the consumer is a spreadsheet, and the Vertex endpoint if you want a dockerised service to call. Within the repository itself, the archived 1.0, 2.0 and 2.5 checkpoints remain available and are worth pinning if you have results you need to reproduce. The broader alternative is the status quo the project argues against — training a bespoke forecasting model per dataset — which still wins when a single problem justifies dedicated attention.
Anyone responsible for forecasting at scale should take this repository seriously. It is a well-documented model with a published paper, an open licence, a live release cadence and, unusually, proof of production use inside the vendor's own products. The honest caution is the support disclaimer: build on the open package for prototypes, internal tooling and evaluation, and move to the hosted surfaces if the forecast ends up on a critical path.