Time series foundation models
Forecast with time series foundation models.
Run forecasts in your browser with TSFM. Compare leading models, explore the research, and find compute providers for your next forecasting task.
Chronos-Bolt Tiny & Mini run on your device · 29 model families to explore
Explore the model landscape
View all modelsTheforecastingcompany
t0
patch Transformer with time and variate attention
102M parameters · Apache-2.0
Prior Labs
TabPFN
in-context tabular Transformer with a time-series regression adapter
parameters · TabPFN-3.5 non-commercial model license
Amazon
Chronos-Bolt
patch-based T5 encoder-decoder with direct multi-step quantile forecasting
9M parameters · See upstream license
Amazon
Chronos / Chronos-T5
T5 encoder-decoder over scaled and quantized time-series tokens
710M parameters · See upstream license
Amazon
Chronos-2
encoder-only, T5-inspired model with group attention and multi-step quantile decoding
120M parameters · See upstream license
TimesFM
Time series transformer with native multivariate and covariate support
parameters · timesfm-non-commercial-license-v1.0 (weights)
Salesforce
Moirai-1.0-R
masked encoder transformer with multi-patch projections, any-variate attention, and mixture-distribution output
14M parameters · See upstream license
Salesforce
Moirai-1.1-R
updated masked encoder Moirai architecture following the 1.0-R design pattern
~14M parameters · See upstream license
Salesforce
Moirai-MoE-1.0-R
sparse MoE decoder-only transformer with probabilistic output heads
~0.47B stored params parameters · CC-BY-NC-4.0
Salesforce
Moirai-2.0-R
decoder-only transformer with quantile loss and multi-token prediction
~11M parameters · See upstream license
NXAI
TiRex
xLSTM-based forecasting model
35M parameters · NX-AI Community License
Datadog
Toto
decoder-only transformer with proportional factorized space-time attention and Student-T mixture output
151M parameters · See upstream license
Choose where your model runs
Run with TSFM
Use Chronos-Bolt Tiny or Mini directly in the TSFM playground. Your browser downloads the model and computes forecasts on your device, keeping your time series local.
Try browser forecastingWith an external provider
Explore managed compute options on model pages, or get weights and deployment tools from Hugging Face. Choose the provider and model that fit your workload.
Find models and providersWhat is a time series foundation model?
Time series foundation models learn from collections of time series and transfer patterns to new forecasting tasks. Many support zero-shot forecasting: applying a pretrained model to a new series without fitting its weights to that series. Some support covariates, multiple related series, quantile forecasts, or other tasks. These capabilities differ across families and versions.
Classical approaches such as ARIMA and Prophet remain useful baselines. Compare models on your actual forecast horizon, data frequency, missing observations, and business error metric. Published benchmark results help you shortlist models, but do not establish accuracy for your own workload.