Chronos-2 is a 120M-parameter, encoder-only time series foundation model designed for robust zero-shot forecasting. Inspired by the T5 encoder architecture, it supports univariate, multivariate, and covariate-informed forecasting tasks within a single unified model. It handles multi-step-ahead probabilistic quantile forecasts and leverages a specialized group attention mechanism for efficient in-context learning across related series and covariates.
Code, tools, and technical resources are publicly available in the official GitHub repository.
Traditional time series foundation models struggle to integrate cross-series interactions and complex external covariates natively. Chronos-2 overcomes these limitations by expanding the classic context window and introducing native support for both historical and future covariate tracks.
Key traits of Chronos-2:
fev-bench, GIFT-Eval, and Chronos Benchmark II.Compared to its predecessors, Chronos-2 expands operational capabilities to cover comprehensive multi-variable workloads:
| Capability | Chronos-2 | Chronos-Bolt | Chronos |
|---|---|---|---|
| Univariate Forecasting | ✅ | ✅ | ✅ |
| Cross-learning across items | ✅ | ❌ | ❌ |
| Multivariate Forecasting | ✅ | ❌ | ❌ |
| Past-only (real/categorical) covariates | ✅ | ❌ | ❌ |
| Known future (real/categorical) covariates | ✅ | 🧩 | 🧩 |
| Max. Context Length | 8192 | 2048 | 512 |
| Max. Prediction Length | 1024 | 64 | 64 |
🧩 Note: Chronos and Chronos-Bolt do not natively support future covariates; they require external regressors that only model per-timestep effects rather than deep trends over time.
Chronos-2 is intended for:
Limitations:
For localized experimentation, you can use the official Python inference library to run zero-shot forecasts directly over PyArrow-backed DataFrames.
pip install "chronos-forecasting>=2.0" "pandas[pyarrow]"
import pandas as pd
from chronos import Chronos2Pipeline
# Load pipeline onto an active hardware device
pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="cuda")
# Load target historical frames and covariate tracks
context_df = pd.read_parquet("https://autogluon.s3.amazonaws.com/datasets/timeseries/electricity_price/train.parquet")
future_df = pd.read_parquet("https://autogluon.s3.amazonaws.com/datasets/timeseries/electricity_price/test.parquet").drop(columns="target")
# Generate probabilistic quantile projections
pred_df = pipeline.predict_df(
context_df,
future_df=future_df,
prediction_length=24,
quantile_levels=[0.1, 0.5, 0.9],
id_column="id",
timestamp_column="timestamp",
target="target",
)
For production-grade environments, the model can be deployed as an elastic cloud microservice.
Provides real-time, serverless, or high-throughput batch operations using clean tabular objects.
pip install autogluon.cloud>=0.5.0
from autogluon.cloud import TimeSeriesFoundationModel
model = TimeSeriesFoundationModel(model_name="chronos-2")
# Option 1: Direct batch prediction
forecast_df = model.predict(df, prediction_length=24)
# Option 2: Spin up a managed real-time inference endpoint
endpoint = model.deploy(instance_type="ml.g5.xlarge")
forecast_df = endpoint.predict(df, prediction_length=24)
Provides granular control over infrastructure definitions using standard JSON request/response payloads.
from sagemaker.jumpstart.model import JumpStartModel
model = JumpStartModel(
model_id="pytorch-forecasting-chronos-2",
instance_type="ml.g5.2xlarge",
)
predictor = model.deploy()
# Send structured payload dict
payload = {
"inputs": [{"target": [1.0, 2.5, 5.3, 12.3]}],
"parameters": {"prediction_length": 24}
}
forecast = predictor.predict(payload)["predictions"]
@article{ansari2025chronos2,
title = {Chronos-2: From Univariate to Universal Forecasting},
author = {Abdul Fatir Ansari and Oleksandr Shchur and Jaris K{\"u}ken and Andreas Auer and Boran Han and Pedro Mercado and Syama Sundar Rangapuram and Huibin Shen and Lorenzo Stella and Xiyuan Zhang and Mononito Goswami and Shubham Kapoor and Danielle C. Maddix and Pablo Guerron and Tony Hu and Junming Yin and Nick Erickson and Prateek Mutalik Desai and Hao Wang and Huzefa Rangwala and George Karypis and Yuyang Wang and Michael Bohlke-Schneider},
year = {2025},
url = {https://arxiv.org/abs/2510.15821}
}