Is Simons state of the art?Pending benchmarks.Announced Nov 11, 2026; no independent results yet.
Abel AI Lab's family of open time-series foundation models for zero-shot quantile forecasting, including with known-future covariates. Four sizes: 25M, 60M, 178M and 297M. Announced November 11, 2026. Public weights, training code, synthetic data generators and public training data are released. No hosted API is offered.
The details so far
- Lab
- Abel AI Lab
- Announced
- Nov 11, 2026
- API
- Not yet · None; the weights and code run locally
- Open weights
- Published · Public, ungated Hugging Face repos abel-lab/Simons-25M, Simons-60M, Simons-178M and Simons-297M, each with model.safetensors
- Parameters
- 25M, 60M, 178M and 297M
- Type
- Family of time-series foundation models for zero-shot probabilistic forecasting, with univariate and multivariate series and known-future covariates (Abel research note and model card)
- Outputs
- Nine quantiles (0.1 through 0.9). Official cards: up to 32,768 time steps of history, 512 steps per forward pass, then longer horizons block by block; up to 32 variates and 16 known-future covariates (Hugging Face model card)
- Architecture
- Each series is scaled and cut into patches of 32 time steps; a causal transformer attends along time while reading known-future covariates one way; a final layer attends across variates (Abel research note)
- Code and data
- Training, inference, synthetic data generators and training recipes at github.com/Abel-ai-lab/Simons. Public training data at abel-lab/Simons-Pretrain; a tech report is described as coming soon (Abel research note)
- License
- Model weights OpenMDW-1.1 (Hugging Face LICENSE files). Training and inference code Apache 2.0 (GitHub LICENSE). The Simons-Pretrain collection is OpenMDW-1.1; each constituent dataset keeps its own license (Abel research note, model cards and NOTICE.md)
- Related
- Separate from Timestral and from the models the lab compares against (TimesFM-3, Toto-2.0 and Chronos-2)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: self-reported results ahead of TimesFM-3, Toto-2.0 and Chronos-2 on GIFT-Eval, fev-bench and TIME, each scored by the benchmark's own measure (Abel research note).
Lab claim: self-reported GIFT-Eval MASE 0.673 and CRPS 0.457, and fev-bench SQL skill score 48.410 and MASE skill score 37.850, all relative to Seasonal Naive, for Simons-297M (Hugging Face model card).
Lab claim: self-reported as running all of GIFT-Eval in under a quarter of an hour on a single A100 GPU, about twenty times faster than models of similar size that take four to five hours, and about ten times faster across GIFT-Eval, fev-bench and TIME (Abel research note).
Lab claim: self-reported that on German day-ahead electricity prices, giving next-day forecasts of demand, wind generation and solar output cuts forecasting error by more than half compared with forecasting from price history alone (Abel research note).
Sources
LABIntroducing Simons (Abel AI Lab)abel.ai/research/simons/ · Nov 11, 2026REPOabel-lab/Simons-297M on Hugging Facehuggingface.co/abel-lab/Simons-297MREPOabel-lab/Simons-Pretrain on Hugging Facehuggingface.co/datasets/abel-lab/Simons-PretrainREPOAbel-ai-lab/Simons on GitHubgithub.com/Abel-ai-lab/SimonsLABAbel AI Lab on X: Simonsx.com/abel_ai_lab/status/2120560598228623468 · Nov 11, 2026LABAbel AI Lab on X: Simons (thread)x.com/abel_ai_lab/status/2120560608869597594 · Nov 11, 2026Tracking continues until an independent board lists it.