Google open-sources TimesFM
Google published TimesFM, a time‑series foundation model trained on ~100B real‑world points that forecasts without dataset‑specific fine‑tuning — a neat tool to plug into end‑to‑end ML pipelines. It’s now public for experimentation and portfolio projects. (x.com)
The Google Research paper "A decoder-only foundation model for time‑series forecasting" was accepted at ICML 2024 and lists Rajat Sen and Yichen Zhou among the authors. (research.google) TimesFM uses a decoder‑only transformer and Google reported a compact ~200 million‑parameter model that achieves strong zero‑shot forecasting performance approaching supervised state‑of‑the‑art on several public benchmarks. (research.google) The project's GitHub notes TimesFM 2.5 as the latest model release, with the 2.5 update reducing the main model to 200M parameters while adding support for up to 16k context length and an optional ~30M‑parameter quantile head for continuous quantile forecasting. (github.com) Official GitHub releases provide PyTorch checkpoints for the 200M model and a v1.2.6 changelog that added support for TimesFM‑2.0 models and related inference fixes. (github.com) Google Cloud’s BigQuery docs show TimesFM exposed inside BigQuery ML via the AI.FORECAST interface and note the model is available in all BigQuery‑supported regions. (docs.cloud.google.com) A Google Cloud blog post describes a larger product framing that cites a pretraining corpus figure of 400 billion real‑world time‑points for TimesFM in the BigQuery integration writeup. (cloud.google.com) The GitHub README explicitly states the open source repo is not an officially supported Google product and the code is released under an Apache‑2.0 license, with model checkpoints mirrored on Hugging Face. (github.com)
