t0-gamma

t0-gamma

Forecasts time series as probabilistic estimates across several horizons, with support for covariates such as holidays and weather. For analysts and developers who test forecasting models.

Made by theforecastingcompany 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: t0-gamma is an open-weights forecasting model from The Forecasting Company. It is a transformer that produces probabilistic forecasts over several horizons and works with extra input series, called covariates. It predicts 21 quantile levels natively, and this is the third public release of the t0 family.

Who it suits: It suits analysts, data scientists and developers who want to test or run time series forecasts on their own data, including forecasts that depend on outside signals.

Industry | Who uses it | How they use it | Example
Energy | Demand analysts | Forecast load with calendar and holiday signals | Electricity demand with a holiday calendar
Healthcare data | Public data analysts | Forecast monthly figures, with and without covariates | Monthly drug reimbursement series
Retail | Planners | Add promotions as known-future covariates | A sales forecast that knows a promotion is planned
Research | ML researchers | Compare open forecasting models on the same data | Run the same series through several open models

The card says the model is meant for research, experimentation and forecast evaluation, so anyone who needs a guaranteed number can skip it.

Where to use it:

• Python 3.11 or newer with PyTorch
• Apple silicon Macs with the MLX runtime, for local inference only
• Retrocast, the maker's hosted site for forecasting your own series
• A managed API, for hosted inference without local weights

Use case | Who | What they get back
Forecast one series | Analysts | A probabilistic forecast over several horizons
Forecast with covariates | Data scientists | A forecast that uses past or known-future signals
Compare forecasts across open models | Researchers | Side-by-side results, which Retrocast supports

How to set it up:

• Python 3.11 or newer
• The tfc-t0 package at version 0.6.1 or later, with one runtime extra
• Disk space for about 1 GB of weights in F32

1. Install the Hugging Face hub client and download the files.
2. Install tfc-t0 with the torch extra, or the mlx extra on Apple silicon.
3. Follow the card's Quickstart for a plain forecast or one with covariates.

pip install -U huggingface_hub
hf download theforecastingcompany/t0-gamma

pip install "tfc-t0[torch]>=0.6.1"   # PyTorch
pip install "tfc-t0[mlx]>=0.6.1"     # MLX, Apple silicon

How to use it: Load the t0-gamma checkpoint in one of the two local runtimes, give it your series, and ask for a forecast. The card's Quickstart covers a plain univariate forecast and a multivariate one that uses past and known-future covariates. A bare install of tfc-t0 adds no runtime, so choose an extra.

Tool | What it does
PyTorch runtime | Runs the checkpoint on broad hardware support
MLX runtime | Runs inference locally on Apple silicon
Managed API | Hosts inference for you, with no local weights
Retrocast | A platform for forecasting your data and comparing open models

Example: You give the model a history of daily electricity demand plus a holiday calendar for the coming weeks. It returns a probabilistic forecast across the horizon, with 21 quantile levels instead of a single line.

For newcomers and experts: A normal forecast gives one number per day. This one gives a range, so you can see how sure it is. Install the torch extra, run the card's Quickstart on a sample series, and then add a covariate such as a holiday flag.

The architecture is a decoder-style patch transformer of about 256M parameters. It handles univariate and multivariate series, historical covariates and known-future covariates, and it forecasts several horizons at once. Known-future covariates can be calendar features, planned events, promotions or weather forecasts. Both local runtimes load the same checkpoint, and the model declares its conventions in config.json.

You provide a series and covariates
Patch transformer reads the history
It predicts 21 quantile levels
A probabilistic forecast comes back

Good to know:

• Forecasts are probabilistic estimates, not guarantees
• Size: about 256M parameters, with the weights near 1 GB in F32
• License: Apache 2.0
• Cost to run: none named for the local runtimes, and the card names no price for the managed API or Retrocast
• Data it reaches: your series files. The managed API and Retrocast run hosted, so data sent there leaves your machine
• Not for: results to treat as certain, since the card calls it a tool for research and evaluation

On Hugging Face, 13 November 2026

• Task: time series forecasting
• Library: tfc-t0
• Parameters: 256M
• Weights: about 1 GB in F32; about 128 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 61,908
• Likes: 32
• License: apache-2.0
• Quantized versions: 1
• Tags: tfc-t0, safetensors, time-series, forecasting, probabilistic-forecasting, foundation-models, pretrained-models, transformer, multivariate, known-future-covariates, open-weights, covariates

Links

• Model: https://huggingface.co/theforecastingcompany/t0-gamma
• Files: https://huggingface.co/theforecastingcompany/t0-gamma/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download theforecastingcompany/t0-gamma

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