Google Research has quietly built what may be the most practical AI model most people have never heard of: TimesFM 2.5, a 200-million-parameter foundation model capable of zero-shot time series forecasting across domains including finance, energy, retail, climate, and healthcare. Available on GitHub under an open licence and integrated into BigQuery ML and Google Sheets, TimesFM represents a significant shift in how predictive analytics is deployed across industries.
What TimesFM Can Do
Time series forecasting — predicting future values based on historical patterns — is one of the most common data science tasks across industries. Retailers need to forecast demand, energy companies need to predict power consumption, banks need to project market movements, and climate scientists need to model weather patterns. Traditionally, each of these use cases required training a separate model on domain-specific data. TimesFM changes this by providing a single pretrained model that can produce accurate forecasts across any domain without additional training, a capability known as zero-shot forecasting.
Key Technical Improvements in Version 2.5
TimesFM 2.5, released in September 2025 and continuously updated through June 2026, represents a significant architectural evolution from version 2.0. The model size dropped from 500 million to 200 million parameters while simultaneously extending context length from 2,048 to 16,384 time points — an eightfold increase. This means the model can analyse longer historical sequences to make more informed predictions. A new continuous quantile head enables probabilistic forecasting (predicting ranges rather than single values), and the removal of the frequency indicator requirement simplifies usage across different data types.
Enterprise Integration and Recent Updates
Google has embedded TimesFM into several enterprise products. BigQuery ML users can run TimesFM-based forecasts using standard SQL queries, and Google Sheets now supports TimesFM-powered forecasting for spreadsheet users. On Vertex AI, the model is available through Model Garden as a Dockerized endpoint. Recent updates include Flax-based inference for faster performance, covariate support through XReg, and LoRA fine-tuning via HuggingFace PEFT, allowing organisations to adapt the model to proprietary datasets with minimal computational cost. In March 2026, the project added agent skill support (SKILL.md), enabling autonomous AI agents to use TimesFM as a forecasting tool.
Benchmark Performance
On the Monash Forecasting Archive — a standard benchmark covering multiple domains — TimesFM outperforms most traditional statistical models like ARIMA and ETS, and matches or exceeds deep learning models trained specifically on the target datasets. On long-horizon tasks such as the ETT datasets, its zero-shot accuracy rivals supervised baselines. A study published in Frontiers in Environmental Chemistry reported a 15–20% reduction in RMSE for energy demand forecasting using TimesFM.
What This Means for India
TimesFM's open-source availability and low computational requirements make it particularly relevant for India's growing AI ecosystem. Indian startups in fintech, agritech, and energy could leverage the model without massive infrastructure investment. The integration with Google Sheets also makes advanced forecasting accessible to small and medium businesses across India — a market where spreadsheet-based workflows remain dominant. For Indian data scientists and ML engineers, TimesFM represents a practical tool that can be deployed quickly for use cases ranging from predicting monsoon patterns to forecasting agricultural commodity prices. The model's 200M-parameter size means it can run on modest hardware, which is an advantage for resource-constrained environments.
