t0-beta ONNX checkpoints

ONNX export of theforecastingcompany/t0-beta for OpenSTEF, produced by openstef-checkpoints.

This export targets a zero-shot time-series forecasting model from the Forecasting Company's t0 family. It is written for a fixed context and horizon window with a dynamic batch axis, so it can be verified against the upstream tfc-t0 reference implementation and published as a portable ONNX checkpoint.

Inputs

Covariates must be point-in-time: the values that existed at the origin, not the values observed afterwards. The graph carries a fixed budget of covariate slots. Fill a slot you have no data for with NaN in both past_covariates and future_covariates and the graph blanks it before the model runs, so any covariate count up to the budget fits this one export.

Export notes

t0 guards every input with an if on a tensor reduction - row types, context width, segment count, LocScale shapes, attention masking, quantile levels - and torch.export cannot trace any of them. Under this contract each guard is statically decidable, so each is replaced by the branch the contract implies. No model mathematics is altered: the transformer, the scaler's statistics and the rollout are the upstream code paths.

Why a re-export and not the first-party ONNX

The Forecasting Company publishes official ONNX exports of t0-beta (t0-beta-onnx-fp16 and t0-beta-onnx-int8). This repository exports its own graph anyway, because the official ones do not fit OpenSTEF's checkpoint contract:

• Different graph interface. The official exports take five inputs (target_context, target_group_ids, future_covariate_context, future_covariate_future, future_covariate_group_ids) under a grouped multivariate convention with no (batch, ...) axis, and require a compute horizon that is a positive multiple of 32, up to 1024. OpenSTEF's forecasters feed past_values, future_covariates and past_covariates as (batch, ...) arrays and read a (batch, horizon, n_quantiles) output, so the official graph is not a drop-in.
• No fp32 variant. Both official artifacts store weights in reduced precision (fp16 / int8) and are tuned for browser, WASM and WebGPU execution. The checkpoints this project publishes are uniform fp32 with a dynamic batch.
• No OpenSTEF metadata. OpenSTEF's runtime reads context length, horizon, covariate count, output name and the native quantile grid from the checkpoint metadata. The official repos ship no such metadata.

A first-party artifact could in principle be wrapped behind an adapter, but that would pin OpenSTEF to a quantized grouped-interface graph and sit outside this repository's trace, verify and publish pipeline. A clean fp32 re-export keeps the interface and precision uniform.

Variants

Export window

• Context: 512 steps
• Horizon: 96 steps
• Resolution: 15 minute step

Provenance

• Source model: theforecastingcompany/t0-beta @ unknown
• Exporter: openstef-checkpoints @ 1e2712e
• Tooling: onnx=1.19.0 onnxruntime=1.27.0 torch=2.14.1 tfc-t0=0.6.0
• Exported: 2026-10-26T20:55:03+00:00

License and attribution

These ONNX checkpoints are derived from theforecastingcompany/t0-beta and released under the same license, apache-2.0. Attribution and all rights to the model weights remain with the upstream authors; if you use the model in research, please cite their work.