Timestral releases four open time-series models, topping GIFT-Eval
Original title: Timestral
Authors say the 209M and 123M variants rank first and second on every metric of GIFT-Eval, TIME, and fev-bench.
Researchers introduce Timestral, four probabilistic time-series foundation models with 5M, 28M, 123M, and 209M parameters. The design uses Decoupled Horizon Forecasting, native known-future covariates in a causal architecture, and pretraining on a large synthetic corpus before mixed real and synthetic post-training. Weights for all four models are on Hugging Face, with inference code on GitHub. The authors report that Timestral-209M and Timestral-123M take the top two places on every metric of GIFT-Eval, TIME, and fev-bench.
Unconfirmed: State-of-the-art rankings are the authors' benchmark claims and are not independently confirmed.
- Model sizes are 5M, 28M, 123M, and 209M parameters.
- Decoupled Horizon Forecasting separates prediction tokens from causally encoded context.
- Pretraining uses only synthetic data; post-training mixes real and synthetic data.
- Weights and inference code are publicly available.
- Organizations
- Hugging Face
From the original article
1 min read at huggingface.co
We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric.
By Adrien Petralia, Alexandre Gavaudan, Artjom Joosen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.