Abel releases small forecasting models with training code and public data
Abel AI Lab announced Simons on November 11, a family of time-series models for forecasting without first fitting a model to each customer's data. Its release includes four sizes, from 25 million to 297 million parameters, alongside training recipes, benchmark code and links to model weights and pretraining data. A GPU is optional for inference. The repository distinguishes Apache-2.0 code, OpenMDW-1.1 model weights and training datasets that retain their own licenses. Simons repository
The release does not expose the entire training corpus: the model card also identifies in-house data. Model card. Its practical attraction is a forecasting pipeline teams can inspect and adapt. Simons returns forecasts at multiple quantiles, providing a range as well as a central estimate. Abel research note
Abel says automated research helped develop the models for less than $8,000 in GPU time. That is a compute claim, not a full project budget. Its benchmark comparisons also need careful reading: GIFT-Eval and fev-bench influenced model selection, and its claimed advantage over TimesFM-3 on one ranking does not extend to the aggregate MASE measure. Independent reproduction was not established in this check. Research methodology
For teams forecasting demand, energy use or other recurring measurements, the release offers a compact starting point for evaluation. Accuracy on their own data remains the deciding test.
Sources and citations
Revision history
Published .