Releases·Model Release
$t_0$: Open-Weights Time-Series Foundation Models for Forecasting with Covariate Context
A new arXiv report introduces $t_0$, a family of open-weights time-series foundation models that condition forecasts on past covariates and known-future covariates without task-specific retraining. The first two releases, t0-alpha (102M) and t0-beta (256M), rank third on GIFT-Eval among zero-shot TSFMs and perform strongly on fev-bench and an independent Macrocosm evaluation.
0 engagement1 sourceMon, Sep 21, 2026, 01:22 PM
AI Overview
- Authors release t0-alpha (102M) and t0-beta (256M), open-weights forecasters conditioning on target history, past covariates, and known-future covariates without retraining.
- Transformer layers alternate attention along time and across variates; forecasts are probabilistic via quantile predictions.
- Pretraining mixes curated public data with synthetic generator families built to contain covariate-to-target dependencies.
- On GIFT-Eval, t0-alpha scores 0.4941 CRPS and t0-beta 0.4738 CRPS / 0.6865 MASE, third on both, within 4.0% of the best zero-shot TSFM.
- On fev-bench the models score 42.2 and 46.7 skill, t0-beta third and 2.0 points behind the leader; known-future covariates raise t0-alpha skill by 6.3 points across 30 tasks.
- t0-beta ranks among the most accurate on the Victoria electricity-demand benchmark with nearly a year of context.
- In Macrocosm's independent evaluation of hourly ERCOT prices over 29 months, both models cut the lagged-price baseline MAE by 38%.
Related
Developments·Paper
Community analyzes t0-alpha, a 101.6M-parameter foundation model for multivariate time-series forecasting
A Reddit user detailed the architecture of t0-alpha, a 101.6M-parameter transformer foundation model for multivariate time-series forecasting. The model separates time attention (temporal evolution) from group attention (cross-variable information exchange), using 24 transformer blocks (16 time-attention, 8 group-attention) and patch embedding of 32 time steps into 512-dimensional representations.
A Reddit user detailed the architecture of t0-alpha, a 101.6M-parameter transformer foundation model for multivariate time-series forecasting. The model separates time attention (temporal evolution) from group attention (cross-variable information exchange), using 24 transformer blocks (16 time-attention, 8 group-attention) and patch embedding of 32 time steps into 512-dimensional representations.
0 engagement1 sourceredditTue, Jul 14, 2026, 06:36 PM
Releases·Model Release
Show HN: T0: We trained a 256M SOTA open-weights time-series forecasting model
The Forecasting Company launched t0-gamma, a 256M-parameter open-weights foundation model for zero-shot time-series forecasting. It ranks first among time-series foundation models on GIFT-Eval, fev-bench and TIME, and runs on CPU, GPU or Apple silicon. Weights are on Hugging Face, with a pip-installable package, Colab notebook, and arXiv tech report.
The Forecasting Company launched t0-gamma, a 256M-parameter open-weights foundation model for zero-shot time-series forecasting. It ranks first among time-series foundation models on GIFT-Eval, fev-bench and TIME, and runs on CPU, GPU or Apple silicon. Weights are on Hugging Face, with a pip-installable package, Colab notebook, and arXiv tech report.
10 engagement1 sourcehackernewsWed, Oct 14, 2026, 02:51 PM
Releases·Model Release
Mistral releases Timestral time series foundation models for zero-shot forecasting in 5M–209M sizes
Mistral released Timestral, a family of time series foundation models for zero-shot forecasting, in 5M, 28M, 123M, and 209M parameter checkpoints on Hugging Face. The models support univariate and multivariate forecasting with covariates and probabilistic predictions across 21 quantiles, with source code released on GitHub for reproducing benchmark results.
Mistral released Timestral, a family of time series foundation models for zero-shot forecasting, in 5M, 28M, 123M, and 209M parameter checkpoints on Hugging Face. The models support univariate and multivariate forecasting with covariates and probabilistic predictions across 21 quantiles, with source code released on GitHub for reproducing benchmark results.
122 engagement2 sourcesreddit, rssThu, Oct 15, 2026, 03:32 PM
Developments·Paper
OpenTSLM TeeMoE: LoRA Time-Series Language Model Unifies Forecasting, Contextual Prediction, and Reasoning
Researchers introduce OpenTSLM TeeMoE, a generalist time-series language model that forecasts from raw series, reasons over textual context, and aggregates external numerical specialists. It trains three LoRA experts over a shared backbone with a learned mixture-of-experts controller, ranking top-three on GIFT-Eval, Context is Key, and TimeSeriesExam.
Researchers introduce OpenTSLM TeeMoE, a generalist time-series language model that forecasts from raw series, reasons over textual context, and aggregates external numerical specialists. It trains three LoRA experts over a shared backbone with a learned mixture-of-experts controller, ranking top-three on GIFT-Eval, Context is Key, and TimeSeriesExam.
0 engagement1 sourcearxivWed, Sep 30, 2026, 05:43 PM
Developments·Paper
Benchmark Finds Zero-Shot Time Series Foundation Models Trade Point Accuracy Against Probabilistic Calibration
An arXiv benchmark study evaluates six Time Series Foundation Models on energy, traffic, and financial datasets against statistical baselines and a supervised deep learning model. TSFMs win on accuracy but face a trade-off between point accuracy and probabilistic calibration. The results offer deployment guidance on balancing generalization with uncertainty quantification.
An arXiv benchmark study evaluates six Time Series Foundation Models on energy, traffic, and financial datasets against statistical baselines and a supervised deep learning model. TSFMs win on accuracy but face a trade-off between point accuracy and probabilistic calibration. The results offer deployment guidance on balancing generalization with uncertainty quantification.
0 engagement1 sourcearxivTue, Sep 22, 2026, 07:21 AM