Chronos-2

Chronos-2

Forecasts time series without any training on your data, covering single series, many related series and series with extra inputs. A 120M-parameter model from Amazon that runs on GPU or CPU.

Made by amazon on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Chronos-2 is a time series foundation model from Amazon that forecasts without training on your own data. One model handles a single series, several linked series and series with extra inputs such as prices or weather. It gives forecasts as quantiles, so you see a range rather than one number.

Who it suits: Analysts and developers who forecast numbers over time and want a ready model instead of building one per series. It suits teams with many related series, since the model learns from them together.

Industry | Who uses it | How they use it | Example
Energy | Market analysts | Forecast prices or load with extra inputs | Predict electricity prices 24 steps ahead, as the card's own example does
Retail | Planning teams | Forecast demand across many related products | Predict sales for a group of items from past sales and promotions
Finance | Data scientists | Produce forecast ranges for planning | Read the 0.1, 0.5 and 0.9 quantiles for a series
Software | Developers | Add forecasting to an app with a few lines of Python | Pass a pandas DataFrame in and get forecasts back

You can skip it if you need to explain why a forecast moved. The model gives predictions, not reasons.

Where to use it:

• Python on Windows, macOS and Linux, with the chronos-forecasting package
• A single GPU or a CPU, since the card supports both
• Amazon SageMaker, through AutoGluon-Cloud or JumpStart, for production

Use case | Who | What they get back
Zero-shot forecast | Analysts | Quantile forecasts for the series you pass in
Forecast with covariates | Data scientists | Predictions that use past and known future inputs
Production serving | Platform teams | Real-time, serverless or batch inference on SageMaker

How to set it up: You need Python and the chronos-forecasting package.

• Python on Windows, macOS or Linux
• The chronos-forecasting package, version 2.0 or newer
• A GPU is optional, and the card's example selects cuda

1. Install Python if you do not have it.
2. Install the package with pip.
3. Load the pipeline and pass your data.
4. For production, deploy to SageMaker with AutoGluon-Cloud.

pip install "chronos-forecasting>=2.0"

pip install autogluon.cloud>=0.5.0

How to use it: Load the pipeline from the model name, then call predict_df on a pandas DataFrame of history. The DataFrame holds an id column, a timestamp column and the target values. Add a second DataFrame with known future covariates if you have one.

from chronos import Chronos2Pipeline

pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="cuda")

pred_df = pipeline.predict_df(
    context_df,
    future_df=future_df,
    prediction_length=24,  # Number of steps to forecast
    quantile_levels=[0.1, 0.5, 0.9],  # Quantiles for probabilistic forecast
    id_column="id",  # Column identifying different time series
    timestamp_column="timestamp",  # Column with datetime information
    target="target",  # Column(s) with time series values to predict
)

Example: You have hourly electricity prices and the known future values of a covariate. You ask for the next 24 steps at the 0.1, 0.5 and 0.9 quantiles. You get a table with a low, middle and high forecast for each step.

For newcomers and experts: A time series is a list of numbers over time, such as daily sales. This model reads your past numbers and predicts the next ones with no training. A first small step is to run the quickstart pattern on one series and look at the three quantile columns.

It is an encoder-only model inspired by the T5 encoder. A group attention mechanism lets it learn across related series and covariates in context. On the fev-bench, GIFT-Eval and Chronos Benchmark II tests, the card puts its zero-shot accuracy at the top among public models. One A10G GPU reportedly handles more than 300 forecasts each second.

Capability | Chronos-2 | Chronos-Bolt | Chronos
Univariate forecasting | Yes | Yes | Yes
Learning across items | Yes | No | No
Multivariate forecasting | Yes | No | No
Past-only covariates | Yes | No | No
Known future covariates | Yes | Needs add-ons | Needs add-ons

Pass a DataFrame of past values
Add known future covariates if any
The model reads related series together
Get quantile forecasts per step

Title: Maximum context length in steps
Chronos-2: 8192 steps
Chronos-Bolt: 2048 steps
Chronos: 512 steps

Good to know:

• The maximum prediction length is 1024 steps, against 64 for Chronos-Bolt and Chronos.
• The older models can use external covariate regressors, which model effects per time step only.
• The weights are about 478 MB in F32, with 119M parameters.
• License: apache-2.0.
• Cost to run: none named locally. A SageMaker deployment needs an AWS account, and the card gives no prices.
• Data it reaches: the series you pass in, plus files downloaded from Hugging Face.
• Not for: explaining the causes behind a forecast.

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Library: chronos-forecasting
• Parameters: 119M
• Weights: about 478 MB in F32; about 60 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 8,097,695
• Likes: 575
• License: apache-2.0
• Quantized versions: 10
• Fine-tunes: 19
• Tags: chronos-forecasting, safetensors, t5, forecasting, time-series-forecasting

Links

• Model: https://huggingface.co/amazon/chronos-2
• Files: https://huggingface.co/amazon/chronos-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download amazon/chronos-2

Then follow the usage in its model card. Nothing in this kit runs by itself.