TimesFM 2.5 200M

TimesFM 2.5 200M is Google's smaller and more deployment-friendly open TimesFM checkpoint. The official repo describes it as an upstream 16K-context model with an optional continuous quantile head, no frequency indicator requirement, and speed-oriented structural updates such as QKV fusion. It halves the parameter count of the 2.0 500M release while extending the usable history window, which is what makes it practical to serve at lower latency without dropping back to a tiny model.

Architecturally it is a decoder-only patched transformer: history is split into patches, encoded autoregressively, and decoded into a forecast horizon. The base model produces point forecasts; an optional 30M continuous quantile head adds calibrated probabilistic outputs when you need prediction intervals rather than a single trajectory. Unlike earlier TimesFM checkpoints it does not require a frequency indicator, so you can feed irregular or mixed-cadence series without hand-labeling their sampling rate. Pretraining draws on GiftEvalPretrain, Wikimedia Pageviews through November 2023, Google Trends through end-2022, and a synthetic/augmented mixture, giving it broad zero-shot coverage across domains.

On TSFM.ai it is served with a 1,024-point hosted context budget — the compiled runtime configuration, not the upstream 16K theoretical maximum — and supports point forecasting, quantile forecasting, and covariates out of the box. Reach for it when you want a strong general-purpose zero-shot forecaster that is cheaper and faster than the 500M model but still long-context enough for daily and sub-daily series; step up to TimesFM 2.0 500M only when a specific workload rewards the larger capacity.