TimesFM-3 is the newest generation of the TimesFM time-series foundation models, natively pre-trained for multivariate forecasting. The model contains 330 million parameters and was pre-trained on a real-world and synthetic time-series corpus comprising more than 1 trillion time points. Since TimesFM's debut in 2024, time-series foundation models have been applied across domains such as retail, finance, observability, manufacturing, healthcare and the natural sciences. Until TimesFM-2.5 (released in September 2025) the family was limited to univariate forecasting; TimesFM-3 extends that capability to complex multivariate scenarios out of the box.
Why multivariate matters
Most real-world forecasting problems depend on multiple interacting time series and auxiliary features. For example, forecasting ice cream sales benefits from related product sales, historical foot traffic, and known future signals like weather forecasts, promotions and holidays. TimesFM-3 supports these patterns natively:
- Multiple targets: jointly forecast several related time series (with both point and quantile forecasts for each target).
- Past covariates: include features known only historically (e.g., past foot traffic).
- Past–future (dynamic) covariates: incorporate known future events (e.g., planned promotions or weather forecasts).
Architecture and inference
TimesFM-3 builds on the decoder-only transformer architecture used by its predecessors. Time series are processed efficiently by grouping contiguous data points into patches of 32 time steps. Per-time-series normalization (similar to TimesFM-2.5) is applied to handle widely differing scales across series.
Multivariate token construction: for target and past-covariate series a token is constructed directly from a single patch. For past–future covariates, TimesFM-3 uses a "lookahead" strategy: each token concatenates the current patch with future patches so the model can access known upcoming signals.
Alternating attention architecture: tokenized patches pass through an input residual block into a transformer stack that operates as a 2D grid with two alternating attention mechanisms:
- Causal temporal attention: tokens attend horizontally across time but only to past tokens within the same time series to prevent data leakage.
- Full variate attention: tokens attend vertically across series at a given time step, allowing the model to learn cross-series correlations (for instance, how a promotion in one series affects another).
These two attention types alternate over several layers, blending temporal patterns with cross-series relationships.
Non-autoregressive decoding: full-horizon prediction in one pass
Unlike earlier versions that generated forecasts patch-by-patch, TimesFM-3 uses Contiguous Patch Masking to produce the entire forecasting horizon in a single forward pass. The model appends masked placeholder tokens for the future horizon alongside the observed context. Target and past-covariate series are masked in the horizon (their future values are unknown), while past–future covariates remain visible so known future signals can guide the forecast. Through the alternating attention layers the model fills in all masked horizon patches simultaneously, with no iterative loop required. For each target series and horizon step the model predicts nine quantiles (from the 10th to the 90th percentile), delivering a full probabilistic view of forecast uncertainty.
Illustrative example: ice cream sales and promotions
If you want to forecast next month’s sales given a planned promotion schedule, a univariate model will project historical weekly patterns forward but cannot anticipate planned promotions. TimesFM-3’s multivariate mode takes the promotion schedule as a past–future covariate, learns the historical relationship between promotions and sales lift, and applies that to future promotion days. In the example provided by the authors this leads to roughly a 20% sales bump on promotion days; the model’s quantile forecasts also reflect uncertainty around those lifts.
Evaluation and benchmarks
TimesFM-3 was evaluated on three public forecasting benchmarks: Gift-Eval, FEV-Bench and Time. On all three, TimesFM-3 achieved the top rank among pre-trained foundation models for both point and probabilistic forecasting metrics. The authors report two evaluation modes: an univariate mode (no covariates or cross-series information) and a full multivariate mode. Even in univariate mode TimesFM-3 matches or exceeds competing models, and in multivariate mode it attains the best average rank across both point and probabilistic measures.
Availability and roadmap
TimesFM-3 is available on GitHub and Hugging Face. A BigQuery integration is scheduled to arrive in the coming weeks. Meanwhile, TimesFM-2.5 remains immediately available for univariate tasks and can be tried via the AI.FORECAST command in BigQuery without ML expertise.
Acknowledgements
The project is joint work with Yichen Zhou, Petros Mol, Abhimanyu Das and Samet Oymak.
Quick technical summary
- Parameters: 330 million
- Pre-training corpus: >1 trillion time points
- Patch size: 32 time steps
- Quantiles predicted: 9 (10th to 90th percentile)
- Supported inputs: multiple targets, past covariates, past–future covariates
- Decoding: Contiguous Patch Masking, non-autoregressive
- Benchmarks: Gift-Eval, FEV-Bench, Time
- Availability: GitHub, Hugging Face; BigQuery integration forthcoming
This article summarizes the TimesFM-3 announcement and technical details based on the information provided by the authors.
