Google's time-series model shrinks, drops frequency hints entirely

What it does

TimesFM is a decoder-only foundation model for time-series forecasting. You feed it historical sequences—plain numpy arrays—and it returns point forecasts and optional continuous quantile predictions up to a 1,000-step horizon. The latest 2.5 release loads like a standard HuggingFace checkpoint and runs in PyTorch or Flax.

The interesting bit

The model shrank from 500M to 200M parameters while expanding its context window to 16k and, notably, eliminating the frequency indicator previous versions required. Google is already embedding it in BigQuery ML and Google Sheets, yet the open repo carries an explicit disclaimer that it is not an officially supported product.

Key highlights

• 200M-parameter architecture with a 16k context length and an optional 30M quantile head for continuous quantile forecasts.
• Covariate support via XReg and no required frequency indicator, suggesting the model generalizes seasonality internally.
• Dual backends: PyTorch for familiarity, Flax for faster inference.
• Fine-tuning pipeline via HuggingFace Transformers and PEFT (LoRA) included in the examples.
• Already deployed inside Google’s own stack: BigQuery ML, Google Sheets, and Vertex Model Garden.

Caveats

• The open release is explicitly “not an officially supported Google product.”
• Legacy 1.0 and 2.0 code is archived in a subdirectory, so the repo carries historical baggage.

Verdict

Try it if you want off-the-shelf forecasting without pipeline boilerplate. If you need a vendor SLA, stick to the Google Cloud integrations instead.