Google open‑sources TimesFM
Google published TimesFM, a foundation model for time‑series forecasting that’s zero‑shot ready and claims to outperform custom models for stock prices, sales and energy demand. The release could speed internal forecasting and give firms a new baseline model for demand and pricing scenarios. (x.com)
Google Research published the TimesFM code and model checkpoints on GitHub and the Hugging Face Hub under an Apache‑2.0 license, with the GitHub repo showing roughly 11k stars and ~929 forks. (github.com) Public checkpoints available on the Hub include timesfm‑1.0‑200m and timesfm‑2.5‑200m, while Google Cloud’s BigQuery product runs a larger TimesFM 2.0 variant reported at 500 million parameters. (huggingface.co) Google Research states the model’s original pretraining corpus was roughly 100 billion real‑world time‑points, while Google Cloud’s BigQuery blog describes a TimesFM variant pretrained on about 400 billion time‑points for its managed preview. (research.google) The open 2.5 checkpoint is described as a 200M‑parameter model that supports long contexts (up to 16,384 points) and an optional ~30M continuous quantile head for probabilistic forecasts. (huggingface.co) BigQuery exposes TimesFM through the AI.FORECAST function and says the native integration can forecast “millions” of univariate series in minutes with a single SQL query as part of the TimesFM preview. (cloud.google.com) The public repo and model cards warn the Lingvo dependency does not support ARM/Apple Silicon, the v1 docs advise (in their v1 subtree) a minimum machine setup for inference, and the GitHub release notes that the open repo is not an “officially supported Google product” while BigQuery provides official managed support. (huggingface.co)
