Amazon Chronos: Time Series Forecasting Models
Updated: March 16, 2026
Complete technical guide to Amazon's Chronos family — T5-based models for zero-shot probabilistic time series forecasting (8M to 710M parameters)
Important Correction: There Is No “Chronos-70B”
Several websites (including a previous version of this page) described a “Chronos-70B” — a 70-billion parameter time series forecasting model. This model does not exist.
The real Amazon Chronos models are based on T5 architecture and range from 8 million to 710 million parameters. The largest model, Chronos-Large, has 710M parameters — roughly 100x smaller than the fabricated “70B” claim.
This page has been rewritten with accurate information from the original Chronos paper (arXiv:2403.07815) and the official GitHub repository.
The Real Chronos Model Family
Chronos is a family of pretrained probabilistic time series forecasting models developed by Amazon Science. Published in March 2024, the paper “Chronos: Learning the Language of Time Series” (Ansari et al., 2024) introduced a novel approach: tokenize real-valued time series into discrete bins, then train standard language models (T5) on these tokens using cross-entropy loss.
Key Innovation
Chronos treats time series forecasting as a token prediction problem. It scales and quantizes time series values into a fixed vocabulary of bins (e.g., 4096 tokens), then uses standard language model training. At inference, it samples multiple future trajectories to produce probabilistic forecasts with uncertainty estimates — no task-specific fine-tuning required.
Training Data
Trained on a large corpus combining: (1) publicly available time series datasets from domains like energy, transport, retail, and weather, and (2) synthetic data from Gaussian processes (TSMix and KernelSynth augmentation). This mix enables strong zero-shot generalization to unseen domains and data patterns.
Technical Specifications
How Chronos Works
Scaling & Quantization
Raw time series values are first normalized (mean scaling), then quantized into one of 4096 discrete bin tokens. This converts continuous values into a discrete vocabulary that T5 can process, similar to how text tokenizers convert words into token IDs.
Token Prediction (T5)
The tokenized historical values are fed into the T5 encoder. The decoder then autoregressively predicts the next tokens (future time steps), using standard cross-entropy loss — exactly like language modeling, but over time series bins instead of words.
Probabilistic Sampling
At inference, Chronos samples multiple future trajectories (e.g., 20 samples) using temperature-based sampling. These samples are dequantized back to real values. The collection of trajectories provides a probabilistic forecast — you get median predictions plus uncertainty bands (e.g., 10th/90th percentiles).
Model Sizes & Hardware Requirements
All five Chronos models are available on HuggingFace. Even the largest (710M) is small enough to run on consumer hardware. Source: HuggingFace Chronos Collection
| Model | HuggingFace ID | Parameters | Download Size | RAM Needed | Best For |
|---|---|---|---|---|---|
| Chronos-Tiny | amazon/chronos-t5-tiny | 8M | ~33MB | ~200MB | Edge devices, quick prototyping |
| Chronos-Mini | amazon/chronos-t5-mini | 20M | ~80MB | ~400MB | Balanced speed/accuracy |
| Chronos-Small | amazon/chronos-t5-small | 46M | ~180MB | ~800MB | Good general forecasting |
| Chronos-Base | amazon/chronos-t5-base | 200M | ~760MB | ~2GB | Production workloads |
| Chronos-Large | amazon/chronos-t5-large | 710M | ~2.8GB | ~4GB | Best accuracy, still runs on laptop |
Key takeaway: Unlike LLMs that need 16-48GB+ VRAM, Chronos-Large runs comfortably on a laptop with 4GB RAM. No GPU required (though GPU speeds up batch inference). This makes Chronos one of the most accessible AI forecasting tools available.
Minimum Hardware for Each Size
System Requirements
Performance Benchmarks
Zero-Shot Performance on Benchmark Datasets
From the Chronos paper (arXiv:2403.07815), aggregated across 27 benchmark datasets. Chronos models were not trained on any of these datasets — this is zero-shot performance. Lower WQL (Weighted Quantile Loss) is better. The paper reports Chronos-Large as competitive with models that were specifically trained on each dataset.
Relative Forecast Quality (higher = better, normalized from paper WQL)
Note: Scores normalized for visualization. See the paper for exact WQL/MASE numbers per dataset. Original metrics are losses (lower = better) — inverted here for intuitive display.
What Chronos-Large Beats (Zero-Shot)
From Table 2 in the paper, aggregate across 27 datasets:
- > Seasonal Naive baseline
- > AutoETS (exponential smoothing)
- > AutoARIMA
- > AutoTheta
- = DeepAR (task-specific trained)
- = PatchTST (task-specific trained)
Key Benchmark Findings
- Zero-shot generalization: Chronos-Large matches or beats models that were trained specifically on each benchmark dataset
- Synthetic data helps: Models trained with KernelSynth augmentation outperform those trained only on real data
- Scale matters: Larger Chronos models consistently outperform smaller ones, following typical scaling laws
- Probabilistic calibration: Prediction intervals from Chronos are well-calibrated across most datasets
Installation & Setup
Install the Chronos package
Install from PyPI (includes all dependencies)
Or install from GitHub (latest)
For the newest features and Chronos-Bolt support
Verify installation
Quick check that everything works
Download a model (automatic on first use)
Models download from HuggingFace automatically. Or pre-download:
Code Examples
Basic Forecasting with ChronosPipeline
This is the real API from the official repository:
Batch Forecasting Multiple Series
Common Mistakes to Avoid
- Wrong:
from transformers import AutoModelForTimeSeriesForecasting— This class does not exist in HuggingFace Transformers. - Wrong:
amazon/chronos-70b— This model ID does not exist. Useamazon/chronos-t5-largefor the largest model. - Right:
from chronos import ChronosPipeline— Use the officialchronos-forecastingpackage. - Right: Pass
torch.tensorfor context, get back numpy-convertible predictions.
Chronos vs Alternatives
| Model | Size | RAM Required | Speed | Quality | Cost/Month |
|---|---|---|---|---|---|
| Chronos-Large | 710M | 4GB | Medium | 85% | Apache 2.0 |
| Chronos-Small | 46M | 1GB | Fast | 73% | Apache 2.0 |
| TimesFM (Google) | 200M | 2GB | Fast | 82% | Apache 2.0 |
| Prophet (Meta) | N/A | <1GB | Fast | 65% | MIT |
| AutoARIMA | N/A | <1GB | Fast | 60% | Open Source |
Quality scores are approximate relative rankings based on the Chronos paper benchmarks. Prophet and ARIMA require per-series fitting; Chronos and TimesFM work zero-shot.
Choose Chronos When
- You need forecasts for many diverse series without per-series model tuning
- You want probabilistic predictions with uncertainty estimates
- You have limited historical data (zero-shot works with short series)
- You need a local, offline solution that runs on standard hardware
Consider Alternatives When
- Multivariate required: Chronos is univariate — use PatchTST or iTransformer for correlated series
- Very long horizons: For 1000+ step predictions, specialized models may be better
- Interpretability needed: ARIMA/ETS provide explicit decomposition
- Real-time streaming: Chronos adds model-loading overhead; lighter methods may suit better
Chronos-Bolt (Newer)
Released in late 2024, Chronos-Bolt models offer up to 20x faster inference than original Chronos while maintaining comparable accuracy.
- Available as
amazon/chronos-bolt-smalletc. - Same
ChronosPipelineAPI - Better for production batch workloads
Limitations & When Not to Use
Univariate Only
Chronos processes one time series at a time. It cannot model cross-variable dependencies (e.g., temperature affecting energy demand). For multivariate forecasting, consider PatchTST, iTransformer, or TimesFM.
No Exogenous Variables
You cannot pass external covariates (holidays, promotions, weather) as inputs. The model forecasts purely from the historical pattern of the target variable itself.
Context Length Ceiling
Default context is 512 time steps. Very long history (thousands of data points) must be truncated, which may lose early patterns. Chronos-Bolt partially addresses this.
Not a Financial Oracle
Chronos cannot predict stock prices, crypto movements, or other financial instruments with actionable accuracy. No time series model can reliably predict efficient markets. Use Chronos for demand, energy, weather, and other domains where historical patterns are genuinely predictive.
Frequently Asked Questions
Does Chronos-70B exist?
No. There is no 70-billion parameter Chronos model. Amazon's Chronos family includes five models based on T5 architecture: Chronos-Tiny (8M), Chronos-Mini (20M), Chronos-Small (46M), Chronos-Base (200M), and Chronos-Large (710M). The largest model has 710 million parameters — not 70 billion. Some websites fabricated a 'Chronos-70B' that does not exist.
What hardware do I need to run Chronos models locally?
Chronos models are lightweight. Chronos-Large (710M) runs on any modern laptop with 4GB+ RAM. The smaller models (Tiny through Small) can run on devices with just 2GB RAM. GPU acceleration helps but is not required — even CPU inference is practical for these model sizes.
How does Chronos compare to ARIMA and Prophet?
Chronos offers zero-shot forecasting — it works on new time series without training, unlike ARIMA which must be fitted per series. The Chronos paper (arXiv:2403.07815) shows Chronos-Large is competitive with or outperforms statistical baselines (seasonal naive, ETS, AutoARIMA) and deep learning models (DeepAR, PatchTST) across 27 benchmark datasets, while requiring no task-specific training.
What is the difference between Chronos and Chronos-Bolt?
Chronos-Bolt (released late 2024) is a faster variant that uses a different architecture optimized for speed. While original Chronos uses T5 encoder-decoder, Bolt models provide up to 20x faster inference. Both are available on HuggingFace under the amazon namespace.
Can Chronos handle multivariate time series?
The original Chronos models are designed for univariate time series — they process one variable at a time. For multivariate forecasting, you would run separate Chronos predictions per variable or use covariates as additional context. Specialized multivariate models like TimesFM or PatchTST may be better suited for highly correlated multivariate data.
Sources
- Ansari, A. F., et al. (2024). “Chronos: Learning the Language of Time Series.” arXiv:2403.07815
- Amazon Science. “Chronos Forecasting.” GitHub Repository
- HuggingFace. “Amazon Chronos Models Collection.” HuggingFace
Was this helpful?
Related Guides
Continue your local AI journey with these comprehensive guides
Written by Pattanaik Ramswarup
AI Engineer & Dataset Architect | Creator of the 77,000 Training Dataset
I've personally trained over 50 AI models from scratch and spent 2,000+ hours optimizing local AI deployments. My 77K dataset project revolutionized how businesses approach AI training. Every guide on this site is based on real hands-on experience, not theory. I test everything on my own hardware before writing about it.