256M-parameter open-weights model t0 tops time-series forecasting benchmarks

LTMeyer · hn · 2026-10-14

Release: The Forecasting Company launched t0-gamma, a 256M-parameter open-weights time-series foundation model (TSFM) that runs on CPU, GPU, or Apple silicon, with weights on Hugging Face, a Colab demo, and an arXiv tech report.
Capabilities: Zero-shot forecasting for nearly any domain or frequency, picking up weak signals like trends and seasonality, and leveraging future-known inputs such as holidays, promotions, and weather. Under the hood it's a Transformer alternating attention across time and variates.
Results: Cuts error by a third vs. a naive repeat-yesterday baseline on European day-ahead electricity prices; ranks first among TSFMs on GIFT-Eval, fev-bench, and TIME, beating ARIMA/ETS, with independent validation on ERCOT prices.
Usage: Install tfc-t0[torch] and load with a few lines of Python; can also plug into agentic, routing, or ensembling systems for further gains.

More from Models