Moirai-2.0-R-Small

Moirai-2.0-R-Small

Forecasts future values of a time series with a small decoder-only transformer trained on many kinds of series. A Salesforce research model for analysts and data scientists.

Made by Salesforce 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: Moirai 2.0 R Small is a time series forecasting model from Salesforce with 11.4 million parameters. It is a decoder-only transformer meant to forecast many kinds of series without per-dataset training. Salesforce released it for research purposes only.

Who it suits: It suits analysts, data scientists and researchers who forecast numbers over time and want a small pretrained model to try.

Industry | Who uses it | How they use it | Example
Retail | Demand analysts | Forecast a sales series from its past values | Project the next weeks of daily orders
Energy | Grid and load analysts | Forecast usage over time | Project hourly demand
Research | Time series researchers | Benchmark a small pretrained forecaster | Compare it with older models on public data

Anyone who needs a model for business use or for high-risk decisions can skip it, since the release is for research only.

Where to use it:

• Python, through the uni2ts library from Salesforce
• A Jupyter notebook, since the library ships a simple example
• A shell where you can make a virtual environment
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Forecast a series | Data scientists | Predicted values for the steps that follow
Series with gaps | Analysts | A model whose patch embedding carries missing-value information
Research comparison | Researchers | A small 11.4M parameter reference to test against

How to set it up: You need a few things first.

• Python and pip
• Git, to clone the uni2ts repository from Salesforce's GitHub
• The virtualenv tool

1. Clone the uni2ts repository, then open its folder.
2. Create and activate a virtual environment.
3. Build the library from source with the notebook extras.
4. Create an empty .env file.

cd uni2ts
virtualenv venv
. venv/bin/activate
pip install -e '.[notebook]'
touch .env

To fetch the model files on their own, use the Hugging Face tool.

pip install -U huggingface_hub
hf download Salesforce/moirai-2.0-R-small

How to use it: The quickest start is the simple example notebook in the uni2ts repository, which runs a forecast with Moirai. You open it in Jupyter and follow its cells with your own series.

Tool | What it does
uni2ts | The Salesforce library that runs Moirai 2.0 inference
moirai_forecast.ipynb | The simple example notebook for getting started
.env file | A file you create in setup, which the library expects
hf download | Fetches the model files from Hugging Face

Example: You open the example notebook with a series of past values, such as daily sales. Moirai 2.0 reads that history and returns a forecast for the steps that follow, which you can plot next to the real values.

For newcomers and experts: A forecasting model looks at numbers recorded over time and estimates what comes next. For a first try, set up uni2ts, open the example notebook and run it as it is before swapping in your own data.

Compared with the first Moirai, this version uses a quantile loss instead of a distributional one, and predicts several tokens at once for better efficiency and stability. It filters out series that cannot be forecast, adds missing-value information to the patch embedding, and masks random patches so inference holds up better. Pretraining mixed GIFT-Eval data, Chronos-derived data, KernelSynth series and internal Salesforce operational data.

Past values are cut into patch tokens
Patches carry missing-value information
Decoder-only transformer reads the history
Several future tokens predicted at once
Forecast is returned

Good to know:

• License: cc-by-nc-4.0, a Creative Commons licence for non-commercial use.
• Cost to run: no API key or paid service named. The model is about 46 MB in F32 by the record.
• Data it reaches: the series you give it. Setup clones a repository and installs packages over the network.
• Not for: business use or high-risk decisions. The card asks you to check accuracy, safety and fairness before deploying anything built on it.

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Parameters: 11.4M
• Weights: about 46 MB in F32; about 6 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 262,191
• Likes: 53
• License: cc-by-nc-4.0
• Quantized versions: 1
• Fine-tunes: 1
• Tags: safetensors, forecasting, time-series, time-series-forecasting

Links

• Model: https://huggingface.co/Salesforce/moirai-2.0-R-small
• Files: https://huggingface.co/Salesforce/moirai-2.0-R-small/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Salesforce/moirai-2.0-R-small

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