Lag-Llama

Definition

Lag-Llama is an open-source foundation model for probabilistic time-series forecasting, built on a decoder-only transformer architecture and pretrained on a large collection of publicly available time series. It is designed to produce zero-shot forecasts for univariate time series and to be freely fine-tuned or extended by researchers, in contrast to forecasting models that are offered only as closed, API-gated commercial services.

Overview

Much of the early progress in time-series foundation models came from proprietary or API-gated systems, which limited how freely academic and independent researchers could inspect, modify, or build on the underlying architecture. Lag-Llama was released as an open-weights alternative aimed at making probabilistic forecasting research reproducible, using a fully published training recipe and model checkpoint that anyone can download and run. Mechanically, Lag-Llama adapts the decoder-only transformer architecture popularized by large language models like Llama to numeric time-series input, using lagged values of the series as input features alongside a learned covariate encoding, rather than the value-discretization approach used by some peers. It is trained with a probabilistic forecasting objective that outputs parameters of a predictive distribution at each future time step, which lets it generate not just a point forecast but a full range of plausible future trajectories through sampling. Compared with other time-series foundation models such as Chronos, Moirai, and TimeGPT, Lag-Llama's main distinction is its fully open training and release process paired with a simpler, univariate-focused design; it does not natively handle multivariate series or variable-frequency inputs with the same built-in flexibility as Moirai. This makes it a lighter-weight option suited to research and experimentation rather than a general-purpose production forecasting service. In practice, researchers use Lag-Llama as a baseline for evaluating new forecasting architectures, as a starting checkpoint for domain-specific fine-tuning, or as a teaching example for how transformer architectures can be repurposed for numeric sequence modeling outside of natural language. Its open license and small footprint relative to API-based alternatives make it practical to run in academic compute environments. Being an early, community-driven open-source project, Lag-Llama's pretraining corpus and model scale are smaller than those behind some commercially backed alternatives, which can translate into weaker zero-shot accuracy on domains underrepresented in its training data. It also lacks native multivariate support, so series with important cross-variable dependencies may be better served by a model designed for that case, and, as with any zero-shot forecaster, results should be validated against domain-specific baselines before production use. The project's community-driven maintenance model also means updates, bug fixes, and new feature support may arrive more slowly and less predictably than with a commercially backed model that has a dedicated engineering team behind it. Documentation and tooling for deploying it in a production pipeline are also less polished than what accompanies a managed API product, so adopters typically invest more integration effort up front.

Key Features

• Fully open-source weights, training recipe, and code release
• Decoder-only transformer architecture adapted from large language models
• Probabilistic forecasting objective producing full predictive distributions
• Zero-shot forecasting for univariate time series
• Uses lagged values and covariate encodings as model input
• Lightweight footprint suited to academic and research compute budgets
• Extensible for fine-tuning on domain-specific forecasting tasks
• Serves as a reproducible baseline for forecasting research

Use Cases

Alternatives

Frequently Asked Questions

Related in AI Models