What it does
TimesFM (Time Series Foundation Model) is a pretrained model from Google Research for time-series forecasting. The repository lets you install it with pip, using a PyTorch backend or an MLX backend for Apple silicon, and includes code examples for running forecasts. The latest version, TimesFM 3.0, adds native multivariate forecasting and flexible covariate support (covariates that cover only the past, or both past and future), and the project reports first place on three forecasting benchmarks (fev-bench, TIME and GIFT-Eval). Forecasts come back as point predictions plus nine quantiles. Earlier versions remain available: 2.5 sits in the main source tree, and 1.0 and 2.0 are archived in a v1 folder. The repository also has a LoRA fine-tuning example, unit tests and an agent skill file. Licensing differs by part. The source code is Apache-2.0, and weights up to version 2.5 are Apache-2.0. TimesFM 3.0 weights use a separate non-commercial license that restricts them to non-commercial, non-production use. Commercial and production use of 3.0 is allowed through authorized Google Cloud services such as BigQuery ML. The page cautions that Google does not officially support this open version.
Capabilities
- Pretrained model for zero-shot time-series forecasting
- Univariate and multivariate forecasting with covariates
- Point forecasts with nine quantiles
- PyTorch backend and MLX backend for Apple silicon
- LoRA fine-tuning example using Hugging Face Transformers and PEFT
When to use it
- You need forecasts for many different series without training a separate model for each.
- You want to forecast several related variables together with extra covariates.
- You want to run forecasting on your own hardware for research or other non-commercial work.
- You need commercial forecasting and are willing to use TimesFM 3.0 through BigQuery ML.
