Lag-Llama
Model directorytime-series-foundation-models/Lag-Llama2.45M params | Apache-2.0 | Upstream checkpoint
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.
Where to run this model
Compute availability, hardware requirements, and pricing depend on the execution provider.
Get model weights, deployment instructions, and current provider listings.
- TimesDB
provider listing · Family / provider variant · Source checked 2026-10-06
The provider documents Lag-Llama prediction and fine-tuning workflows. Account access and endpoint availability are managed by TimesDB.
- sktime TServe
self-hosted runtime · Family / provider variant · Source checked 2026-10-06
Run a model-serving container on your own hardware or compute account. Select the catalog checkpoint and matching runtime image.
Hugging Face provider mappings checked 2026-10-06. No Inference Provider mapping was listed for this checkpoint.
Model Classification
Family
Lag-Llama
Type
time series foundation model
Pretrained time-series model for zero-shot or few-shot forecasting workloads.
Resources
Training Data
Curated subset of public time-series repositories spanning energy, traffic, weather, and economic domains — trained on 27 datasets from the Monash Time Series Forecasting Repository.
Recommended For
- • Probabilistic zero-shot forecasting with well-calibrated uncertainty estimates
- • Lightweight deployment where sample-based distribution quality matters
Strengths
- • One of the earliest open probabilistic TSFMs — well studied and widely benchmarked
- • Lag-based tokenization captures temporal dependencies efficiently with a small model
Limitations
- • Smaller capacity than newer large-scale TSFM families — may plateau on complex multivariate workloads
- • Focused on forecasting rather than multi-task time-series understanding
- • Can under-react on strong deterministic trend extrapolation compared with newer larger zero-shot forecasters
- • Prediction intervals can be very wide with short context — provide at least 100+ data points for tighter uncertainty estimates
Not Ideal For
- • Strong monotonic trend continuation where you expect the first forecast step to stay close to the recent level
- • Using long-context zero-shot quality as the main selection criterion against newer larger model families
Capabilities
Tags
Specifications
- Parameters
- 2.45M
- Architecture
- decoder-only transformer (LLaMA-style) with lag-based tokenization and Student-t mixture output
- Context length
- n/a
- Max context
- n/a
- Minimum history
- n/a
- Recommended history
- n/a
- Input step
- n/a
- Required target series
- 1
- Temperature
- Ignored
- Top P
- Ignored
- Max output
- n/a
- License
- Apache-2.0