Google open‑sources TimesFM
Google released TimesFM, a time‑series AI model pre‑trained on 100 billion data points that can run locally and claims zero‑shot forecasting for sales trends and market prices. The model reportedly outperforms supervised baselines and is being framed as useful for forecasting tasks without heavy task‑specific training. (x.com)
Time-series forecasting is the business of predicting the next numbers in a sequence, from store demand to electricity use. Google has open-sourced TimesFM, a forecasting model its researchers say can be applied to new datasets without task-specific retraining. (research.google) Google Research described TimesFM as a decoder-only model pre-trained on 100 billion real-world time points in its original release, and the paper behind it was presented at the International Conference on Machine Learning in 2024. The code is public in Google Research’s GitHub repository under an Apache-2.0 license. (research.google, github.com, arxiv.org) In plain terms, users give the model a history of past values and ask for the next stretch of values, such as the next week of sales or the next month of traffic. Google’s first open checkpoint, `timesfm-1.0-200m`, was released on Hugging Face for univariate forecasting, meaning one numeric series at a time. (huggingface.co, research.google) Google said the model’s zero-shot forecasts came close to, and in some tests beat, supervised baselines trained separately for individual datasets. The paper reports evaluations across public benchmarks from multiple domains and time granularities rather than a single industry use case. (arxiv.org, research.google) That puts TimesFM in the same broad pattern as large language models, but for numbers ordered over time instead of words ordered in a sentence. Google’s paper says the model uses a patched decoder architecture, which groups time points into chunks so it can process longer histories more efficiently. (arxiv.org) The open-source release also gives Google a wider path into production forecasting tools. Google Cloud said in December 2025 that TimesFM models were being added to BigQuery and AlloyDB, and BigQuery documentation now describes a built-in TimesFM model available in supported regions. (cloud.google.com, docs.cloud.google.com) The repository has continued to move beyond the first public checkpoint. As of April 2026, the GitHub README lists TimesFM 2.5 as the latest model version, with earlier 1.0 and 2.0 versions archived in a `v1` directory. (github.com) Google’s own materials also show the scale has grown since the first paper. The 2023 research blog and 2024 paper cite 100 billion time points, while the December 2025 Google Cloud post says the newer cloud version was pre-trained on more than 400 billion real-world time points. (research.google, arxiv.org, cloud.google.com) The practical pitch is simple: fewer custom models for every forecasting task, and more reuse of one pre-trained model that can run inference from public checkpoints. The harder question, which Google’s paper and docs both imply, is where zero-shot convenience is good enough and where domain-specific tuning still wins. (arxiv.org, github.com, docs.cloud.google.com)
