Google open‑sources TimesFM

Forecasting is the business of guessing the next number in a sequence, from next week’s sales to tomorrow’s power demand. Google Research has released TimesFM as open-source software, letting developers run that kind of prediction with a pre-trained model instead of training one from scratch. (github.com, research.google) Google’s public repository says the latest open version is TimesFM 2.5, and the code is available under the Apache 2.0 license in the `google-research/timesfm` project. The repository README was updated in March 2026 and links to model checkpoints hosted by Google on Hugging Face. (github.com, huggingface.co) TimesFM is a time-series foundation model, which means it is pre-trained on many sequences and then applied to new ones without task-specific retraining. Google Research said the earlier open model was trained on 100 billion real-world time points and was designed for “zero-shot” forecasting on unseen datasets. (research.google, arxiv.org) The core pitch is speed and reuse. Instead of building a separate model for each dataset, a user can load a checkpoint and ask for forecasts directly in Python, and Google Cloud also exposes the model through BigQuery’s `AI.FORECAST` function. (github.com, cloud.google.com) Google’s February 2024 paper described TimesFM as a decoder-only transformer, a neural network architecture adapted from language models to read a history of values and predict the next stretch. That paper was later accepted at the International Conference on Machine Learning, or ICML, in 2024. (research.google, icml.cc) The newest open checkpoint changes some of the practical limits. Google’s README says TimesFM 2.5 uses 200 million parameters, supports context windows up to 16,000 points, and can produce continuous quantile forecasts up to a 1,000-step horizon with an optional 30 million-parameter quantile head. (github.com, huggingface.co) That matters for users working with long histories such as hourly traffic, energy load, or web demand, where older open checkpoints capped context at 512 points for version 1.0 and 2,048 points for version 2.0. Google’s model cards list those earlier limits alongside the newer 2.5 release. (huggingface.co, huggingface.co, github.com) Google has also drawn a line between the research release and the commercial product. The Hugging Face cards say the open checkpoints are “not an officially supported Google product,” while Google Cloud’s documentation says BigQuery’s built-in TimesFM model is the supported implementation for customers. (huggingface.co, cloud.google.com) The performance claims are narrower than some social-media posts suggest. Google’s paper and blog say TimesFM’s zero-shot results came close to state-of-the-art supervised models on several public benchmarks, not that it beat every specialized forecaster on every task. (research.google, arxiv.org) Outside studies have already tested those limits in finance. One October 2024 paper on financial fine-tuning reported that baseline TimesFM performed poorly on direct market price prediction before additional training, while a separate 2024 paper on Value-at-Risk forecasting found the model competitive against some conventional risk models in that narrower setting. (arxiv.org, arxiv.org) So the release is less a promise of one model for every chart than a public toolkit: a Google-built forecaster, open weights, and code that can be tested on new series immediately. Whether it works best out of the box or after adaptation still depends on the data in front of it. (github.com, cloud.google.com, arxiv.org)