t0-gamma brings local forecasting to CPUs, GPUs and Apple silicon
The 256M-parameter open-weight model predicts numerical time series with future-known inputs, but its authors warn that an older runtime can silently degrade results.
Forecast numbers without calling a hosted model
The Forecasting Company is putting t0-gamma in front of developers as a roughly 256M-parameter, open-weight forecasting model that runs on local CPUs, GPUs and Apple silicon. The practical use is numerical prediction: take a history of demand, prices or server metrics and return a distribution of likely future values, rather than a paragraph of advice.
The October 14 Show HN introduction follows the October 13 release of tfc-t0 0.6.0. That package adds support for gamma, the family’s third public model, and unifies its PyTorch and MLX runtimes. Hugging Face records the gamma repository’s creation on October 12, so the community post should not be mistaken for the first appearance of its weights.
The model accepts historical covariates and information known over the forecast horizon, such as holidays, planned promotions or weather forecasts. It predicts 21 quantile levels and can decode up to 1,024 timesteps in one forward pass, using autoregressive rollout for longer horizons.
The version requirement matters
The authors identify a failure that is easy to miss: versions before 0.6.0 can load gamma’s weights without an error while ignoring two numerical conventions in its configuration. They warn that this silently degrades forecasts, especially at longer horizons.
The documented installation is pip install "tfc-t0[torch]>=0.6.0" for PyTorch, or the mlx extra for Apple silicon. Python 3.10 or newer is required; the MLX route also needs macOS 14 or newer and native arm64 Python, and does not install PyTorch. The public checkpoint downloads without authentication. Local inference needs no provider API key, although a separate managed API is available.
Promising scores, with a specific recipe
The model card reports gains over t0-beta on GIFT-Eval, fev-bench and TIME. For example, its reported GIFT-Eval CRPS falls from 0.4738 to 0.4520, where lower is better. These are the team’s reported results, not an AITrending reproduction or a guarantee for a particular business dataset.
The published scores use optional history-based inference settings, including symmetric averaging, which takes two forward passes when enabled and applicable. The authors recommend evaluating calibration, horizon and domain performance on users’ own historical backtests. The weights and runtime code are Apache-2.0 licensed.