A List of Time Series Foundation Models

Definition: Most time-series foundation models are not large language models. They are purpose-built forecasting systems pretrained from scratch on large collections of numerical time series. TIME-LLM is the clearest example here of a model that genuinely repurposes a frozen LLM. The original Chronos is a middle case: it borrows language-model tokenization and T5-family machinery, but it is trained on time-series data rather than natural language.

It’s important to see that difference, especially in 2026. The field has moved from broad claims about “LLMs for forecasting” toward specialized time-series foundation models, stronger zero-shot benchmarks, in-context adaptation, probabilistic forecasting, multivariate inputs, and known-future covariates. LLMs are increasingly used around the forecaster – as context processors, routers, or agentic orchestrators – rather than as the numerical forecasting engine itself.

These labels describe different things:

Chronos-2 is a 120M-parameter, encoder-only foundation model that supports univariate, multivariate, and covariate-informed forecasting. It uses patches and group attention to learn across related series and known covariates in context. This is a different architecture from the original tokenized Chronos-T5 family. The official repository provides code, checkpoints, and deployment examples.

TimesFM 2.5 is the current public release of Google’s decoder-only TSFM. It has 200M parameters, supports context windows up to 16,000 time points, and can produce continuous quantile forecasts up to a 1,000-step horizon with an optional quantile head. Covariates are supported through XReg.

TimesFM-ICF, presented at ICML 2025, is the named in-context fine-tuning method. It continues pretraining so the model can use related example series supplied at inference time, recovering much of the benefit of target-data fine-tuning without updating weights for each new dataset.

Moirai 2.0 replaces Moirai 1.0’s masked encoder, multi-patch inputs, and mixture-distribution outputs with a simpler decoder-only design, a single patch size, quantile forecasting, and multi-token prediction. It was trained on a corpus of 36 million series. The paper reports that it is faster, substantially smaller, and more accurate than the earlier Moirai 1.0-Large comparison model.

The original Moirai 1.0 remains historically important: it was trained on LOTSA, a collection containing more than 27 billion observations, and was designed for cross-frequency and multivariate forecasting. But it should not be presented as the current Moirai architecture.

Toto 2.0 is Datadog’s current recommended release for multivariate forecasting, with five open-weight model sizes from 4M to 2.5B parameters. Its strongest specialization is observability data, and the repository also includes BOOM, Datadog’s benchmark for observability metrics. The technical report describes the scaling recipe and evaluations; the checkpoints and code are released under Apache 2.0.

TiRex is a compact 35M-parameter model based on xLSTM. It provides zero-shot point and quantile forecasts across short and long horizons. TiRex-2 extends the family with native multivariate forecasting and support for past and future-known covariates. This makes “TiRex” a family rather than a single fixed 2025 checkpoint.

TabPFN-TS takes a different route: it converts univariate forecasting into a tabular regression problem, adds seasonal and calendar features, and uses TabPFN as the predictor. Version 1.1 defaults to the TabPFN-TS-3 checkpoint and supports known-future covariates. It is useful when lightweight feature construction and tabular in-context learning fit the problem better than a large autoregressive forecaster.

Sundial is a family of generative TSFMs pretrained on TimeBench, a corpus containing one trillion time points. It produces forecast samples rather than only a point estimate, allowing users to derive means, quantiles, and other statistics for probabilistic zero-shot forecasting. Sundial was accepted as an ICML 2025 oral paper.

Time-MoE is a decoder-only, sparse mixture-of-experts family trained from scratch on Time-300B, an open collection with more than 300 billion time points across multiple domains. The models scale to 2.4B parameters while activating only part of the network for each token. The paper was accepted as an ICLR 2025 spotlight.

TimeGPT is a managed forecasting model accessed through Nixtla’s API. It helped establish the commercial TSFM category, but its weights and training corpus are not public. Claims that it was trained on the “largest publicly available” collection or broadly outperforms traditional methods come from Nixtla’s own paper and should be treated as vendor-reported rather than independently established. The client library is public; the model itself is not.

TIME-LLM genuinely uses a frozen LLM for forecasting. It maps numerical patches toward text prototypes, feeds them through the frozen language-model backbone, and uses natural-language task information through Prompt-as-Prefix.

However, the argument that pretrained language knowledge is itself responsible for better numerical forecasts became much harder to sustain after the NeurIPS 2024 paper “Are Language Models Actually Useful for Time Series Forecasting?” Across three tested LLM-based methods, removing the LLM or replacing it with a basic attention layer did not reduce performance and often improved it. That result does not prove that an LLM can never help forecasting, but it does show that strong ablations are necessary before attributing gains to linguistic pretraining.

MoiraiAgent represents the newer hybrid direction. It uses a lightweight LLM to process context, select among forecasting experts, and decide which preprocessing or post-processing tools to invoke. Its expert pool includes Chronos-2, TimesFM 2.5, and TiRex. Expert selection belongs to MoiraiAgent – not to Moirai 1.0 or 2.0. The available model card describes the release as research-only, so it is better treated as an early agentic framework than as a mature production forecaster.

These papers remain useful, but they should not be mixed into a foundation-model leaderboard without explaining that many are task-specific architectures trained and evaluated differently.

There is no universal winner. Start with the forecasting requirements:

Public benchmark results are a shortlist, not a deployment decision. Evaluate candidates with rolling-origin backtests, several forecast horizons, realistic covariate availability, leakage checks, seasonal-naive and classical baselines, uncertainty calibration, and a loss function aligned with the business cost of errors. Compare model versions explicitly: “TimesFM” or “Moirai” is not precise enough when their architectures have changed.

A time series foundation model is pretrained across many numerical series so it can transfer to new datasets, domains, frequencies, or forecast horizons with little or no task-specific training.

No. Most are purpose-built numerical forecasting models trained on time-series data. Some borrow language-model architecture or tokenization, while systems such as TIME-LLM directly repurpose a pretrained language model.

There is no single best model. Shortlist current versions such as Chronos-2, TimesFM 2.5, Moirai 2.0, Toto 2.0, TiRex-2, TabPFN-TS, Sundial, and Time-MoE, then compare them on your own data and constraints.

Many support zero-shot forecasting. In-context examples, relevant covariates, calibration, or fine-tuning may improve accuracy, but every adaptation must be evaluated without temporal leakage.

Use rolling-origin evaluation, multiple horizons, realistic covariate availability, leakage checks, seasonal and classical baselines, uncertainty calibration, latency, cost, and a metric aligned with the business impact of errors.