Show HN: T0: We trained a 256M SOTA open-weights time-series forecasting model

This project delivers t0, a 256M-parameter open-weights time-series forecasting foundation model that produces probabilistic forecasts from historical data and known-future covariates. Public checkpoints (t0-alpha, t0-beta, t0-gamma) provide out-of-the-box quantile forecasts and median predictions via a simple predict API that accepts NumPy or PyTorch inputs, handles missing values and padding, and supports batching of heterogeneous series through a TimeSeries wrapper and batch_series helper. Worked examples include French national electricity demand, monthly drug reimbursements (Medic'AM) and a European day-ahead electricity price backtest that uses load, renewables, holidays and weather as covariates. The package supplies recommended inference flags to reproduce published scores (trim leading missing, symmetric averaging, clip nonnegative, snap integers) and exposes forward for differentiable passes when fine-tuning is desired.

Architecturally, t0 is a decoder-style patch transformer that alternates time and covariate attention, decoding up to 1,024 future timesteps in parallel and falling back to autoregressive rollout for longer horizons. Quantile levels are interpolated between trained points and extrapolated on exponential tails. Deployment options include PyTorch and an Apple-optimized MLX runtime, plus ONNX artifacts (from t0-alpha) for FP16/INT8 inference and a managed API. The repo includes tooling, notebooks, benchmarks and instructions to install and run locally or at scale (Retrocast).