Time-series foundation models converted to ONNX for Tabular Data Workbench.
The weights in these files are the upstream weights; the graphs around them
were re-exported, so the files are modified forms of the originals (see
"Changes" below).

Chronos-2 and Chronos-2 small
  Source: amazon/chronos-2 (revision 29ec3766d36d6f73f0696f85560a422f50e8498c)
          autogluon/chronos-2-small (revision ddec01313e50b6bc58ebaa92ede81bc24a3d9f9a)
          https://github.com/amazon-science/chronos-forecasting
  Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
  Licensed under the Apache License, Version 2.0 (LICENSE).

Kumo-Anomaly (NV-Tesseract-AD)
  Source: nvidia/Kumo-Anomaly (revision 226c5003b0582adc1cac433123f47324bfae39f2)
          https://github.com/NVIDIA/structured-data-models
  Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  Licensed under the Apache License, Version 2.0 (LICENSE; NVIDIA's copy with
  its copyright line is LICENSE.structured-data-models).
  Kumo-TS includes DPM-Solver, Copyright (c) 2022 Cheng Lu, licensed under the
  MIT License (THIRD-PARTY-NOTICES.txt).

Changes
  - chronos-2*.fp32.onnx: the Chronos-2 network re-exported with
    torch.onnx (dynamo, opset 18). Instance scaling, the arcsinh transform,
    NaN handling and the masks built from group ids are computed outside the
    graph, and the graph contains no int64 shape arithmetic. Weights
    unchanged.
  - kumo-anomaly.fp32.onnx: Kumo-Anomaly's denoiser network re-exported the
    same way. The diffusion sampling loop (DPM-Solver++) and the scoring run
    outside the graph. Weights unchanged.
  - *.f16w.onnx: the same graphs with every weight of 1,024 or more elements
    stored as float16 and cast back to float32 on load.