What Are Time Series Foundation Models?
4 entriesTime series foundation models (TSFMs) are pretrained models that forecast a new series without being trained on it, so one model serves many datasets and domains. They are trained on large collections of real and synthetic series and typically return probabilistic forecasts as quantiles. The Timestral report frames them as prior-data fitted networks, in which forecasting is amortised posterior inference. Classical methods such as ARIMA or Prophet are instead fitted separately to each series.
Timestral (Mistral AI)
4 entriesAccording to the report, Timestral is a family of four probabilistic forecasting models with 5M, 28M, 123M and 209M parameters. The authors report state-of-the-art results on GIFT-Eval, TIME and fev-bench, with the 209M and 123M models taking the top two places on every metric. These are the authors' own claims, so check the benchmark leaderboards before relying on them.
The Models to Know
12 entriesThe Timestral report compares against recent models such as Chronos-2, TimesFM-3, Toto-2.0, TiRex-2, Moirai-2.0, Kairos and FlowState. The entries below link to each project's repository, with release and licence facts as listed by GitHub.
Benchmarks and Evaluation
4 entriesThe Timestral report evaluates on GIFT-Eval, TIME and fev-bench, using error against parameter count to show which models sit on the accuracy-versus-cost frontier. Leaderboards change often, and some entries are self-reported by model authors, so treat rankings as time-bound.
Frequently Asked Questions
What is a time series foundation model?
A pretrained model that forecasts new series without training on them, usually returning probabilistic forecasts as quantiles. One model covers many datasets and domains, unlike ARIMA or Prophet, which are fitted to each series.
What is Timestral?
Timestral is a family of four time series foundation models from Mistral AI, with 5M, 28M, 123M and 209M parameters, described in arXiv:2610.37585 (27 October 2026). The authors report top results on GIFT-Eval, TIME and fev-bench.
What is Decoupled Horizon Forecasting?
Per the Timestral report, it separates prediction tokens from the causally encoded context so longer horizons can be trained without masking observed values, and each observed value serves as both a forecast target and context for later forecasts.
Do TSFMs support covariates?
Some do. Timestral adds past-only and known-future covariates within a causal architecture, and the report notes that Toto-2.0 and Chronos-2 also handle covariates. Check each model's documentation.
Can I use these models commercially?
Licences differ. Chronos, TimesFM, Moirai, Toto and Granite repositories list Apache 2.0, while the Timestral model cards list "other" and GitHub does not identify TiRex or TimeGPT licences. Read the current terms before deploying.
How current is this page?
Links were checked on 30 October 2026. Model facts come from the Timestral report and each project's repository. Benchmark claims are the authors' own and leaderboards change often, so run your own tests on your data.