README.md

metadata

license: openmdw-1.1
library_name: simons
pipeline_tag: time-series-forecasting
datasets:
  - abel-lab/Simons-Pretrain
thumbnail: https://huggingface.co/abel-lab/Simons-297M/resolve/main/assets/abel-color.png
tags:
  - time-series
  - forecasting
  - probabilistic-forecasting
  - foundation-model
  - covariates
  - pretrained
  - pytorch
model-index:
  - name: Simons-297M
    results:
      - task:
          type: time-series-forecasting
        dataset:
          name: GIFT-Eval
          type: Salesforce/GiftEval
        metrics:
          - name: MASE (relative to Seasonal Naive)
            type: mase
            value: 0.6731
          - name: CRPS (relative to Seasonal Naive)
            type: crps
            value: 0.4574
        source:
          name: GIFT-Eval leaderboard
          url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
      - task:
          type: time-series-forecasting
        dataset:
          name: fev-bench
          type: autogluon/fev_datasets
        metrics:
          - name: SQL skill score (%, relative to Seasonal Naive)
            type: sql
            value: 48.41
          - name: MASE skill score (%, relative to Seasonal Naive)
            type: mase
            value: 37.85
        source:
          name: fev-bench leaderboard
          url: https://huggingface.co/spaces/autogluon/fev-bench
      - task:
          type: time-series-forecasting
        dataset:
          name: TIME
          type: Real-TSF/TIME
        metrics:
          - name: MASE (relative to Seasonal Naive)
            type: mase
            value: 0.6371
          - name: CRPS (relative to Seasonal Naive)
            type: crps
            value: 0.5341
        source:
          name: TIME leaderboard
          url: https://huggingface.co/spaces/Real-TSF/TIME-leaderboard

Simons-297M

Simons-297M is an open time series foundation model from Abel AI Lab for probabilistic forecasting. It supports univariate and multivariate series with known-future covariates. The GitHub repository contains inference and training code. The release post, Introducing Simons, tells how the models were built. All four sizes and the training data are in the Simons 1 collection. The Simons Forecast Explorer shows forecasts of the four sizes in the browser.

Model details

• Size: 297.2M parameters.
• Architecture: a causal transformer over patches of 32 time steps, then an attention layer across the variates of a series; 13 layers, width 1280, 20 heads.
• Context and horizon: up to 32,768 time steps of history; 512 steps per forward pass, longer horizons block by block.
• Inputs: up to 32 variates and up to 16 known-future covariates per series, numeric or categorical; NaN marks missing values. Past-only covariates are accepted and dropped.
• Outputs: 9 quantiles (0.1, 0.2, ..., 0.9).
• Training: four stages, from synthetic pretraining to post-training on real data; the released weights are the equal-weight average of the stage-3 model and its continuation (WiSE-FT).
• Inference: PyTorch on CPU or GPU, with the simons package, version 0.1.0 or later.

Benchmarks

The place of Simons-297M and four widely used models on each benchmark. On GIFT-Eval, fev-bench and TIME the places are among the open-weight entries.

Efficiency

Simons is fast to run: on one A100 GPU, Simons-297M forecasts a whole benchmark in minutes, about ten times faster than models of similar size.

Get started

pip install "git+https://github.com/Abel-ai-lab/Simons"

import torch
from simons import Simons

device = "cuda" if torch.cuda.is_available() else "cpu"
model = Simons.from_pretrained("abel-lab/Simons-297M", device=device)

# [batch, time] or [batch, variates, time]; NaN marks missing values
history = torch.randn(8, 2000).cumsum(-1)
quantiles = model.forecast(history, horizon=96)  # [8, 96, 9]

# Known-future covariates: [batch, covariates, time + horizon]
covariates = torch.randn(8, 2, 2000 + 96)
quantiles = model.forecast(history, horizon=96, covariates=covariates)

• Training data: abel-lab/Simons-Pretrain, the public part of the data the models were trained on.
• Getting started: docs/GETTING_STARTED.md covers the input format, covariates, forecast options and fine-tuning.
• Models: docs/MODEL.md compares the four sizes and gives their scores.
• Training: docs/TRAINING.md covers the stages, configs and GPU settings.
• Data: docs/DATA.md describes what the models were trained on.

Training data

• Synthetic series from the open generators in src/simons/synth, with synthetic known-future covariates of known effect.
• Public datasets with a permissive or share-alike license, most of them released as abel-lab/Simons-Pretrain.
• Internal data: an in-house corpus.

See the dataset for data details and the training quickstart for training commands. Notices required by the training data's licenses: NOTICE.md.

License

The weights are licensed under OpenMDW-1.1 (LICENSE); the code on GitHub is licensed under Apache-2.0 (LICENSE). Each training dataset keeps its own license; see NOTICE.md for details.

Citation

@misc{simons2026,
  title  = {Simons: Building Open Time Series Foundation Models through Auto-Research},
  author = {{Abel AI Lab}},
  year   = {2026},
  url    = {https://github.com/Abel-ai-lab/Simons}
}