This repository hosts the Granite Time Series TinyTimeMixer (TTM-R2) model in lightweight, portable formats (GGUF, TorchScript, ONNX, MLX).
The original IBM Granite model is a compact forecasting network designed to handle common time-series patterns: cyclical, random, seasonal, and trend.
The purpose of this repo is to make the model practical for local use on consumer hardware such as Apple Silicon laptops. Instead of needing large GPUs or servers, you can test, benchmark, and run forecasting workflows directly on a MacBook or similar device.
What’s inside
Model formats provided
FP16 GGUF
Quantized GGUF (q6_k)
TorchScript export
ONNX export
MLX (Apple Metal-accelerated format)
Usage examples
Forecasting on synthetic patterns (cyclical, seasonal, trend).
Quick evaluation with provided scripts.
Conversion scripts showing how to move between formats.
Why this matters
Lets you run state-of-the-art time-series forecasting with minimal hardware.
Useful for research, prototyping, or experimentation in finance, IoT, or general prediction tasks.
Provides a base for extending to more complex local forecasting setups.
Audience
Developers and researchers who want small, fast models for forecasting without depending on large GPU clusters.
IBM Granite Time Series Models
Model Card for Granite Time Series Models
Model Details
Model Developers: IBM Research
Architecture: Tiny Time Mixer (TTM)
Input/Output: Multivariate time series → Future value predictions
Granite Time Series models are purpose-built to forecast time series signals. They are designed to capture a wide range of temporal patterns, including:
Cyclical patterns
Random fluctuations
Seasonal effects
Long-term trends
TTM models are lightweight and efficient, making them well-suited for resource-constrained environments and edge devices.
Intended Use
Time-series forecasting for business, industrial, or research use cases.
Evaluation and benchmarking on standard datasets.
Academic exploration of compact time-series architectures.
Limitations
Designed for univariate and multivariate series with typical forecasting horizons.
Not optimized for extremely long sequences or tasks outside forecasting.
Training
Trained on a mixture of synthetic and real-world datasets.
Objective: minimize forecasting error across diverse temporal patterns.
Evaluation
Granite TTM models are benchmarked on common forecasting tasks and compared against classical baselines and other neural architectures. Performance is competitive while maintaining a small footprint.
Ethical Considerations
Ensure proper use in domains where forecasting has real-world consequences (e.g., finance, healthcare).
Misuse of forecasts can lead to poor decision-making; human oversight is required.
Citation
If you use Granite Time Series models in your research, please cite the original IBM Research paper and model card.
@inproceedings{ekambaram2024tinytimemixersttms,
title={Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series},
author={Vijay Ekambaram and Arindam Jati and Pankaj Dayama and Sumanta Mukherjee and Nam H. Nguyen and Wesley M. Gifford and Chandra Reddy and Jayant Kalagnanam},
booktitle={Advances in Neural Information Processing Systems (NeurIPS 2024)},
year={2024},
}
