The Forecasting Company announced the public release of t0‑gamma, a 256 M‑parameter decoder‑style transformer for probabilistic time‑series forecasting. The model can be loaded with a single command:
from t0 import T0Forecaster
model = T0Forecaster.from_pretrained("theforecastingcompany/t0-gamma").eval()
and the required package is installed via pip install "tfc-t0[torch]>=0.6.0".
Quick start and core capabilities
The repo provides a predict method that accepts NumPy or PyTorch tensors, returns quantile forecasts, and handles missing values (NaN) automatically. Users can add known‑future covariates—e.g., weather or holiday calendars—by passing a future_covariates tensor, letting the model condition on future information without forecasting it. Batch inference is supported through the batch_series helper, which aligns series of different lengths and returns a single tensor of shape (batch, horizon, quantiles).
Runtime options and hardware support
Three runtimes are shipped:
- PyTorch (default, works on any CPU/GPU)
- MLX for Apple silicon (pip install "tfc-t0[mlx]>=0.6.0"), which runs on macOS 14+ without needing PyTorch
- ONNX artifacts for t0‑alpha (FP16/INT8) but no ONNX export for t0‑gamma yet, limiting deployment to the two runtimes above.
The model can forecast up to 1024 steps in a single forward pass; longer horizons fall back to autoregressive rollout.
Licensing, provenance, and what’s missing
The code is Apache‑2.0 licensed and builds on ideas from Datadog’s Toto, Amazon’s Chronos‑2, and other open‑source projects. The announcement does not include pricing because the model is free to download, but the managed API is a separate product with undisclosed costs. Moreover, the lack of an ONNX export for t0‑gamma means edge or cross‑platform deployment may require additional engineering effort.
When to try it
If you already use PyTorch or have Apple silicon workstations and need probabilistic forecasts that incorporate future covariates, clone the GitHub repo and run the quick‑start script. Keep an eye on future releases for ONNX support and managed‑API pricing details.
For the full technical description see the accompanying arXiv paper.
