Lag-Llama is an open-source project on GitHub with about 1,601 stars, most recently updated in 2025. Lag-Llama is an open foundation model for probabilistic time series forecasting. Built on a decoder-only transformer architecture inspired by Llama, it uses lagged values of the time series as covariates, allowing a single pretrained model to produce probabilistic forecasts across many different domains and frequencies.
Published on GitHub under the Apache-2.0 license, Lag-Llama has attracted a community of about 1,601 stars, which places it among the more visible open-source AI projects in its niche. Development was still active as of 2025, based on the repository’s latest updates. Because the full source code is public, teams can audit how it works, adapt it to their own stack, and self-host it rather than depending on a closed hosted product.
The repository is tagged with topics including forecasting, foundation-models, lag-llama, llama, time-series, which gives a quick picture of where it fits in the open-source AI ecosystem. New users can explore the examples and documentation in the repository to understand the main workflows before integrating it into their own projects, and the public issue tracker shows what the maintainers are actively working on and what the community is asking for.
Lag-Llama is Apache-2.0 open source and free for research and commercial use under that license.
Lag-Llama is best suited for time series researchers, forecasting engineers and data scientists who want a pretrained foundation model instead of training every forecaster from scratch. Anyone evaluating it should start with the GitHub repository, where the documentation, examples and issue tracker show exactly what the current version can do.