OpenSTEF/chronos-2-onnx

chronos-2 ONNX checkpoints

ONNX export of amazon/chronos-2 for OpenSTEF, produced by openstef-checkpoints.

OpenSTEF resolves these via HubCheckpoint; the metadata file next to each weights file drives the inference path. Each variant was checked against the torch reference on representative inputs before publishing.

Variants

File	Precision	Static shapes	Max deviation vs torch
chronos-2_static.onnx	fp32	yes	2.241e-05
chronos-2.onnx	fp32	no	1.86e-05
chronos-2_int8.onnx	int8	no	0.7023

Pick by deployment target: static fp32 is the portable, CoreML-eligible default; int8 for size; dynamic when context or horizon must vary.

Static and dynamic variants

The dynamic variants leave the input axes free, so the context length, horizon, and number of covariate series can vary at run time. The static variants freeze every axis to a fixed size. That is what makes them eligible for CoreML and quicker to load, at the cost of only accepting the one window they were built for.

The static graph is built for:

• Batch: 4 series (one target plus 3 covariates)
• Context: 5760 steps (60 days at 15-minute resolution)
• Horizon: 672 steps (7 days)

With those sizes its ONNX inputs are fixed to:

Input	Shape
context	(4, 5760)
group_ids	(4,)
attention_mask	(4, 5760)
future_covariates	(4, 672)
future_covariates_mask	(4, 672)

To run a different window, use a dynamic variant.

Provenance

• Source model: amazon/chronos-2 @ unknown
• Exporter: openstef-checkpoints @ 3b897ec44492a3ddc36c5d05b3bac5f33c26d49b
• Tooling: onnx=1.22.0 onnxruntime=1.27.0 torch=2.12.1+cpu
• Exported: 2026-06-19T17:16:49+00:00

License and attribution

These ONNX checkpoints are derived from amazon/chronos-2 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.

Modifications. The weights are not retrained, fine-tuned, or otherwise changed. At the graph level the export does two things, both verified to match the reference model within the deviation shown above: it reimplements a few operators the ONNX tracer does not support (NaN-aware mean and sum, arcsinh, and the patch reshape) with equivalent ONNX ops, and it recomputes the rotary-embedding frequency buffer that the upstream loader leaves uninitialised. It then exports the result in static or dynamic shape variants at fp32 or int8 precision.

The tooling that produced these files (openstef-checkpoints) is licensed MPL-2.0, which does not extend to the weights.