Google has officially launched TimesFM-2.5, a foundational time series model that adopts a pure decoder architecture and is equipped with 200 million parameters. This model can accommodate a context length of up to 16K and is inherently designed with probabilistic forecasting capabilities, thereby propelling time series forecasting closer to real-world practicality. In a zero-shot foundational model evaluation conducted on the GIFT-Eval benchmark, TimesFM-2.5 exhibited remarkable performance. By reducing the parameter count by half and expanding the context length, the model has achieved a simultaneous enhancement in both efficiency and performance. Presently, the model is accessible on the Hugging Face platform, with ongoing efforts to integrate it into BigQuery and Model Garden. This development is poised to expedite the adoption of zero-shot time series forecasting in actual business settings.