Timestral models top time-series benchmarks, weights released
First merged summary.
Researchers released four Timestral probabilistic time-series models, with the 209M and 123M variants claimed first and second on three benchmarks.
Researchers introduced Timestral, a family of 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, synthetic-only pretraining, and mixed real-and-synthetic post-training. The authors report that Timestral-209M and Timestral-123M rank first and second on every metric of GIFT-Eval, TIME, and fev-bench. Weights for all four models are on Hugging Face, and inference code is on GitHub.
- Model sizes are 5M, 28M, 123M, and 209M parameters.
- Authors say 209M and 123M rank first and second on every metric of three benchmarks.
- Benchmarks named are GIFT-Eval, TIME, and fev-bench.
- Pretraining uses only synthetic data; post-training mixes real and synthetic data.
- Weights are on Hugging Face; inference code is on GitHub.
What each outlet adds
beyond the shared facts- Hugging Face daily papersPublic weights, GitHub inference code, and architecture details.
- Organizations
- Hugging Face
Coverage
oldest first · one row per article · UTC- · 4d ago
Hugging Face daily papersFirst report
Timestral releases four open time-series models, topping GIFT-Eval
Authors say the 209M and 123M variants rank first and second on every metric of GIFT-Eval, TIME, and fev-bench.
- · 3d ago
arXiv cs.AI / cs.LG / cs.CL
Timestral models lead GIFT-Eval, TIME and fev-bench
Four probabilistic forecasters from 5M to 209M parameters ship with Hugging Face weights and GitHub inference code.