Author: Dr. Andreas Fehlner, ONNX Steering Committee
The ONNX (Open Neural Network Exchange) community is excited to announce the release of ONNX v1.21.0, a significant step forward in advancing open standards for machine learning interoperability. This release introduces Opset 26, along with new operators and precision types that enable richer, more flexible model representations for modern AI workloads. Notably, support for 2-bit data types and the addition of CumProd and BitCast operators expand what developers can express in ONNX graphs, while continued investments in security and tooling reinforce the project’s commitment to a robust and trustworthy ecosystem.
Key Highlights
- Opset 26 introduced
- A new version of ONNX’s operator standard, enabling AI models to express more functionality and run across a broader range of tools and runtimes.
- 2-bit data type support
- Models can now use ultra-compact 2-bit representations, improving efficiency for edge, mobile, and embedded deployments.
- New operators: CumProd and BitCast
- CumProd: Enables cumulative multiplication across tensors
- BitCast: Allows reinterpretation of data without copying
- These additions give developers more flexibility in defining model behavior.
- Improved consistency in integer division
- Resolves a long-standing ambiguity, ensuring more predictable and consistent model execution across frameworks.
- Enhanced model migration support
- Expanded version conversion helpers make it easier to upgrade older models to newer ONNX standards.
- Support for the latest Python advancements
- Experimental compatibility with Python 3.14’s free-threading mode helps future-proof the ecosystem.
- Strengthened security and build tooling
- New compiler hardening options align with industry best practices, improving safety for production use.
Read the full release notes here.
What’s Next
The ONNX community is already looking ahead to the next release cycle, with a focus on expanding operator support for generative AI workloads, improving quantization capabilities, and extending version converter coverage. A newly formed Probabilistic Programming Working Group is also underway, aiming to bring Bayesian inference and probabilistic modeling into ONNX as first-class capabilities.
Now is a great time to get involved and help shape the future of the project. Contributions of all kinds are welcome, from new operators and tooling improvements to documentation and testing.
We’d also love to hear from you — share your thoughts in our Community Survey.
To get started, visit onnx.ai, explore the project on GitHub, or join the ONNX community meetings.
