Google's 200M-parameter time-series foundation model with 16k context
Summary
TimesFM is a decoder-only time-series foundation model from Google Research optimized for forecasting, now with 200M parameters and a 16k context window in its 2.5 release. The update adds covariate support, an enhanced inference API, and broader ecosystem integration (Hugging Face, BigQuery), while remaining open-source though not an official Google product. The article provides release notes, installation guidance, and example code to experiment with timesfm both in PyTorch and Flax/JAX backends.