Chronos-Bolt Small

Chronos-Bolt Small is a pretrained time series forecasting model that uses a T5 encoder-decoder architecture to generate zero-shot probabilistic forecasts. It processes historical time series data by chunking observations into patches, which the encoder analyzes, while the decoder directly produces multi-step quantile forecasts. The model was trained on nearly 100 billion time series observations.

Key Strengths

Chronos-Bolt models are more accurate, up to 250 times faster, and 20 times more memory-efficient than the original Chronos models of the same size. The Small variant has 48 million parameters. In benchmark evaluations across 27 datasets, zero-shot Chronos-Bolt models outperform both commonly used statistical models and deep learning models that were trained on those datasets (marked with * in the benchmark plot). They also surpass other foundation models that had prior exposure to benchmark data (marked with +). Notably, Chronos-Bolt Base exceeds the original Chronos Large in forecasting accuracy while being over 600 times faster.

Benchmark Performance

The following plots compare inference speed and accuracy (Weighted Quantile Loss and Mean Absolute Scaled Error) across 27 datasets from the Chronos paper.

Model Variants

Model	Parameters	Architecture Base
chronos-bolt-tiny	9M	t5-efficient-tiny
chronos-bolt-mini	21M	t5-efficient-mini
chronos-bolt-small	48M	t5-efficient-small
chronos-bolt-base	205M	t5-efficient-base

Chronos models have been downloaded over 120 million times from Hugging Face. The model supports both CPU and GPU inference and is licensed under Apache-2.0.