ONNX
ONNX is an open-source project that defines a standard format for representing machine learning models, enabling interoperability between different AI frameworks and hardware platforms.
Profile
ONNX provides an open file format and operator specification that lets AI models move between different training frameworks and deployment runtimes without rewriting code.
ONNX is an open-source project that defines a standard format for representing machine learning models, enabling interoperability between different AI frameworks and hardware platforms. Founded in 2017 and hosted under the Linux Foundation, it is not a company and has no employees, revenue, or venture funding. The project's GitHub repository (github.com/onnx/onnx) shows 20,900 stars, 3,900 forks, and 3,809 commits as of June 2026, with active development led by a community of contributors including engineers from Microsoft, Facebook (Meta), Amazon, and other founding partners.
ONNX provides a common set of operators and a file format (.onnx) that allows models trained in PyTorch, TensorFlow, or other frameworks to be exported and run on different runtimes (e.g., ONNX Runtime, TensorRT) and hardware (CPUs, GPUs, specialized accelerators). The project is governed by a steering committee and includes Special Interest Groups (SIGs) for areas like operators, converters, and performance. While ONNX has no direct revenue or customers, its specification is widely adopted: major cloud providers (AWS, Azure, GCP) and hardware vendors (NVIDIA, Intel, AMD) offer ONNX support, and the format is used by thousands of organizations for model deployment.
The project's main challenge is maintaining backward compatibility and operator coverage as new model architectures (e.g., transformers, diffusion models) emerge faster than the standard can be updated. As of mid-2026, the ONNX community is working on ONNX 2.0, which aims to simplify the operator set and improve support for dynamic shapes and control flow. There are no recent funding rounds, acquisitions, or layoffs because ONNX is not a legal entity.
Who buys this
- AI/ML engineers who need to deploy models across multiple hardware platforms (CPU, GPU, NPU)
- Cloud platform teams building inference services that must support models from various frameworks
- Hardware vendors (chipmakers, accelerator designers) who want their devices to run models from any framework
- Enterprise data science teams that standardize on one training framework but deploy on heterogeneous infrastructure
- Open-source tool developers building model converters, optimizers, or runtime engines
Strengths and what to watch
Strengths
- Widely adopted industry standard: supported by all major cloud providers (AWS, Azure, GCP) and hardware vendors (NVIDIA, Intel, AMD, Qualcomm), reducing vendor lock-in for model deployment.
- Large, active open-source community: 20,900 GitHub stars, 3,900 forks, and 3,809 commits as of June 2026, with contributions from Microsoft, Meta, Amazon, and others, ensuring ongoing maintenance and operator coverage.
- Broad framework support: ONNX converters exist for PyTorch, TensorFlow, Keras, Scikit-learn, and many other frameworks, allowing models to be exported without significant code changes.
Watch for
- Operator coverage lag: New model architectures (e.g., state-space models, diffusion transformers) often require custom operators that are not yet in the ONNX standard, forcing users to fall back to framework-specific runtimes.
- Governance and contributor concentration: While nominally community-governed, the majority of core commits come from a small number of large companies (Microsoft, Meta), raising questions about long-term neutrality and responsiveness to non-sponsor needs.
- Backward compatibility burden: The ONNX specification must maintain backward compatibility across versions, which slows the addition of new features and can lead to a bloated operator set that is hard to optimize for hardware vendors.
Key Information
- Industry
- AI Frameworks, Tools & Libraries
- Founded
- 2017
Frequently Asked Questions
What is ONNX and what does it do?
ONNX is an open-source project that defines a standard format for representing machine learning models. It allows AI models to move between different training frameworks and deployment runtimes without rewriting code, enabling interoperability across various hardware platforms.
Who created ONNX and who maintains it?
ONNX was founded in 2017 and is hosted under the Linux Foundation. It is not a company and has no employees or funding. A community of contributors, including engineers from Microsoft, Meta, and Amazon, leads development, governed by a steering committee and Special Interest Groups.
Which AI frameworks and hardware support ONNX?
ONNX converters exist for PyTorch, TensorFlow, Keras, and Scikit-learn. Major cloud providers like AWS, Azure, and GCP, along with hardware vendors such as NVIDIA, Intel, and AMD, offer ONNX support, allowing models to run on CPUs, GPUs, and specialized accelerators.
What are the main benefits of using ONNX for model deployment?
ONNX reduces vendor lock-in by providing a common format for models trained in different frameworks. It is widely adopted across cloud and hardware platforms, enabling deployment on heterogeneous infrastructure without rewriting code, and has a large active community ensuring ongoing support.
What are the limitations or challenges of ONNX?
New model architectures like transformers and diffusion models often require custom operators not yet in the ONNX standard, forcing fallback to framework-specific runtimes. Maintaining backward compatibility slows feature additions and can bloat the operator set, making optimization harder for hardware vendors.
What is ONNX 2.0 and when is it coming?
As of mid-2026, the ONNX community is working on ONNX 2.0, which aims to simplify the operator set and improve support for dynamic shapes and control flow. This update addresses challenges from emerging model architectures and seeks to enhance performance and compatibility.
Sources
- github.com — GitHub repository statistics (stars, forks, commits), project structure, and recent commit history as of June 2026.
- tracxn.com — Tracxn profile confirming ONNX is unfunded, founded in 2017, based in Menlo Park, and operates as an open standard for ML interoperability.
- news.crunchbase.com — Context on Q1 2026 AI venture funding landscape, showing $300B invested globally, with 80% going to AI companies; ONNX is not mentioned as a recipient.
- forgeglobal.com — List of top Q1 2026 funding rounds (OpenAI, Anthropic, xAI, Databricks); ONNX is absent, consistent with its unfunded status.