Open Neural Network eXchange

Description

Open Neural Network eXchange (ONNX) is an open-source, vendor-neutral format for representing machine learning (ML) and deep learning models. In the context of 3GPP Release 19, ONNX is adopted as a standardized model representation format to facilitate AI/ML model exchange, distribution, and lifecycle management across 5G-Advanced network functions. The specification work is captured in 3GPP TR 26.847 (Study on AI/ML model transfer and format) and TR 26.927 (AI/ML media and network integration), which define how ONNX models are packaged, transported, and executed within the 3GPP ecosystem.

Architecturally, ONNX sits at the model representation layer of the 3GPP AI/ML pipeline. A trained model is serialized into an ONNX graph consisting of nodes (operators), tensors (data), and metadata (model version, input/output signatures, operator set version). This graph is then transported between network functions — for example, from an ML training function to an ML inference function — using 3GPP-defined interfaces. The receiving entity can parse the ONNX graph, validate operator compatibility, and execute inference using an ONNX runtime or a vendor-specific runtime that supports the ONNX operator set.

Key components of the ONNX-based model exchange framework include: the ONNX model file itself (typically .onnx), the operator set (opset) version that defines available operators and their semantics, model metadata describing input/output tensors and preprocessing requirements, and optional auxiliary files such as tokenizers or feature extractors. 3GPP specifications define how these components are bundled, how model identity and versioning are managed, and how model updates are distributed to inference endpoints without service interruption.

ONNX plays a critical role in enabling multi-vendor AI/ML interoperability in 5G-Advanced networks. Without a common model format, each vendor would need to re-train or convert models for its own runtime, leading to fragmentation and increased integration cost. By standardizing on ONNX, 3GPP allows an ML model trained by one vendor to be deployed on another vendor's inference engine, provided both support the required ONNX operator set. This is particularly important for use cases such as network energy saving, load balancing, mobility optimization, and quality-of-experience prediction, where models may be trained centrally and deployed across many network elements.

In terms of network integration, ONNX models are treated as managed artifacts within the 3GPP management framework. They have identifiers, versions, and lifecycle states (training, validation, deployment, retirement). The management system can query model capabilities, trigger model updates, and monitor inference performance. Security considerations include model integrity protection, authenticity verification, and protection against adversarial inputs — aspects addressed in 3GPP security specifications for AI/ML. Overall, ONNX in Rel-19 represents a foundational enabler for interoperable, lifecycle-managed AI/ML in 5G-Advanced and future 6G networks.

Purpose & Motivation

ONNX exists to solve the problem of AI/ML model interoperability across heterogeneous network functions and vendors. Before its adoption in 3GPP, AI/ML models in telecom networks were typically vendor-proprietary, tied to specific runtimes and hardware. This created significant integration barriers: a model trained by one vendor could not easily be deployed on another vendor's inference engine, and network operators had limited flexibility in sourcing and deploying AI capabilities. The lack of a common format also hindered model lifecycle management, as each vendor used its own packaging, versioning, and metadata conventions.

The motivation for standardizing on ONNX in 3GPP Release 19 stems from the growing role of AI/ML in 5G-Advanced network optimization. Use cases such as network energy saving, mobility optimization, load balancing, and QoS prediction require models to be trained, distributed, and executed across multiple network elements. A standardized model format enables operators to train models centrally, validate them once, and deploy them broadly across multi-vendor infrastructure. It also facilitates model sharing between operators and vendors, and supports regulatory requirements for model transparency and auditability.

Historically, previous approaches relied on proprietary model formats or ad-hoc conversion tools, which were error-prone and did not scale. ONNX, already widely adopted in the broader AI industry, offered a mature, well-specified, and tool-supported alternative. By adopting ONNX, 3GPP leverages existing industry momentum while adding telecom-specific requirements for model transport, security, and lifecycle management. This approach reduces development cost, accelerates time-to-market for AI features, and ensures that 3GPP AI/ML remains aligned with mainstream AI tooling.