Modular-TS-2.4M

A 2.4M-parameter time-series foundation model, pretrained from scratch by Modular Computing.

Modular-TS-2.4M is a small, fast, zero-shot probabilistic forecaster: give it any univariate or multivariate series and it returns 21 quantiles for every future step, with no fine-tuning. It is built for the GPU-poor: it runs comfortably on a CPU or a phone-class device.

Results (GIFT-Eval, full 97 configurations, zero-shot)

Geometric mean relative to Seasonal Naive (lower is better).

Model	Params	MASE	CRPS
Modular-TS-2.4M (ours)	2.41M	0.745	0.511
IBM TTM-R3 (PT)	~1–5M	0.724	0.520
TimeSSM-2.5M	2.5M	0.754	0.513
Seasonal Naive	–	1.000	1.000

• Best CRPS (probabilistic accuracy) of any open model at 2.5M parameters or fewer that we evaluated.
• Beats TimeSSM-2.5M on both metrics and TTM-R3 on CRPS; TTM-R3 is still ahead on MASE.
• Our numbers are strictly zero-shot: the model never sees a test dataset's own training split at inference time. (TTM-R3-PT uses each dataset's train split for backtest ensembling.)
• Evaluated with the official GIFT-Eval harness at leaderboard settings; the same harness reproduces published leaderboard scores for reference models exactly.

Model

Architecture	Decoder-style patch transformer (Timestral-style), d_model 128, 2 heads, MLP 320, 8 layers alternating time and variate attention, patch 32
Parameters	2,407,680 (embed 107,392 / trunk 1,804,672 / head 495,616)
Context	up to 8,192 steps
Output	21 monotone quantiles per step (plus mean)

The architecture code adapts layers from Mistral AI's Apache-2.0 timestral inference package. No Timestral weights were loaded, fine-tuned, or distilled at any point. Every weight in this repo was trained from random initialisation.

Training data

• Synthetic (majority): a GPU-vectorised generator with Gaussian-process kernels, regime-switching Ornstein–Uhlenbeck processes, waveform and load-curve templates, steps, growth curves, intermittent counts, multivariate lead-lag mixing, covariates and missingness, followed by an augmentation cascade.
• Real (~40% in the final stage): GiftEvalPretrain, resolution-sampled with TSMixup, after a contamination filter that removes any dataset overlapping the GIFT-Eval test sets.

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("Modularcomputing/Modular-TS-2.4M")
# path/export/model.safetensors + path/export/config.json
# Load with the bundled modular_ts package (see modular_ts/README.md),
# which wraps the Apache-2.0 timestral forecaster around these weights.

Files:

• export/model.safetensors, export/config.json: inference weights and config
• modular_ts/: training, synthetic data generation and evaluation code
• results/all_results.csv: per-configuration GIFT-Eval scores

Licence

• Weights: CC BY-NC 4.0 (research and non-commercial use), because the real-data portion of training mixes datasets under several licences.
• Code (modular_ts/): Apache-2.0, including the adapted timestral components (see modular_ts/NOTICE).

Limitations

• Trails TTM-R3 on point accuracy (MASE); weakest on some long-horizon, high-frequency energy and weather series.
• A small model: it captures common seasonal and trend structure well but can miss rare regime changes.

Citation

@misc{modularts2026,
  title  = {Modular-TS-2.4M: a small time-series foundation model trained from scratch},
  author = {Akuegbo, Chima Emmanuel and Modular Computing},
  year   = {2026},
  url    = {https://huggingface.co/Modularcomputing/Modular-TS-2.4M}
}