Lag-Llama

Lag-Llama is an open probabilistic forecasting model released through the Time Series Foundation Models org by ServiceNow Research. As one of the earliest open probabilistic TSFMs, it is widely studied and benchmarked, and at 2.45M parameters it remains one of the smallest public TSFMs with native uncertainty output.

Architecturally it is a LLaMA-style decoder-only transformer that uses lag-based tokenization and a Student-t mixture output head. It converts lagged historical values into tokens and autoregressively generates full predictive distributions, an efficient way to capture temporal dependencies in a very small model. It is trained on a curated subset of public time-series repositories spanning energy, traffic, weather, and economic domains, drawing on 27 datasets from the Monash Time Series Forecasting Repository, and is released under Apache-2.0.

On TSFM.ai reach for Lag-Llama as a lightweight probabilistic baseline where well-calibrated, sample-based distribution quality matters more than peak point accuracy. The upstream best-practices guide recommends sweeping context length starting from the 32-point training regime, and intervals can be wide with short history, so provide at least 100+ points for tighter uncertainty. It currently suits probabilistic baseline usage better than aggressive deterministic trend continuation, where its smaller capacity can under-react relative to newer, larger zero-shot forecasters.