Menu
Latest
12 meetings · Prague, Jeju · 12–16 Oct
Reports
RAN1#126-bis · SA2#177

Glossary term · Management

AI/ML model data

Artificial Intelligence / Machine Learning Model Data

Management →

AI/ML model data is the internal components of a trained neural network, including architecture, weights, biases, and parameters, that define input processing and model outputs and are exchanged and managed in 3GPP AI/ML frameworks.

Introduced
Rel-19
Where
Services › Codecs
Specifications
2 specs
AI/ML model data Description Purpose What changed Neighbours Specifications

Description

AI/ML model data refers to the internal constituents of a trained neural network that collectively determine the mapping from inputs to outputs. As defined in 3GPP specifications TS 26.114 and TS 26.927, model data includes the model architecture (the arrangement of layers, neurons, and connections), the weights and biases learned during training, and other parameters such as activation functions, normalization statistics, and hyperparameters that influence inference behavior. In essence, model data is the information that must be present for a model to perform inference; without it, the model is merely a structural description.

In the 3GPP AI/ML management architecture, model data is a managed artifact that can be produced, stored, transferred, and consumed by various functions. A training function (for example, an AI/ML training function in OAM or a model training capability in the NWDAF) generates model data as the output of training. This model data is then packaged and delivered to an inference function, which loads it to perform predictions or analytics. The management framework defines services for distributing model data, monitoring its performance, and updating it when retraining occurs. Model data may also be exchanged between federation members during federated learning, where intermediate models carry model data snapshots.

Model data is distinct from training data and inference data. Training data is the input used to learn the parameters, inference data is the input presented at runtime, and model data is the learned artifact itself. This separation is fundamental to AI/ML lifecycle management: it allows model data to be versioned, validated, and deployed independently of the data it was trained on. In 3GPP contexts such as the NWDAF, model data may be associated with specific analytics types (for example, load analytics or mobility analytics), and its provenance, accuracy, and validity period are tracked as metadata.

The representation and transfer of model data raise practical considerations addressed by 3GPP. Model data can be large, particularly for deep neural networks, so specifications consider compression, format standardization, and efficient delivery mechanisms. Security is also critical: model data may be proprietary and may encode sensitive information about the training data, so integrity protection and access control are required. TS 26.927, for example, addresses AI/ML media formats and related model data handling, reflecting the intersection of AI/ML with media services.

Model data also plays a role in interoperability and multi-vendor deployments. By standardizing what constitutes model data and how it is described, 3GPP enables a training function from one vendor to deliver model data to an inference function from another. This is essential for building open, composable AI/ML pipelines in 5G and 6G networks, where analytics and automation are increasingly central to network operation and service delivery.

Purpose & Motivation

As 5G networks incorporated AI/ML for analytics, automation, and optimization, 3GPP needed a precise, standardized notion of what constitutes a trained model so that training and inference functions could interoperate. Earlier specifications focused on analytics outputs and inputs but did not formally define the internal artifact — the model itself — that produces those outputs. This gap made it difficult to manage model lifecycle, transfer models between functions, and ensure multi-vendor compatibility.

The AI/ML model data concept was created to fill this gap. By defining model data as the architecture, weights, biases, and parameters of a trained neural network, 3GPP provides a common vocabulary for model management. This enables management services to distribute, version, validate, and monitor models in a standardized way. It also supports the separation of concerns between training (which produces model data) and inference (which consumes it), allowing each to be implemented and deployed independently.

Furthermore, the concept supports privacy and security objectives. Because model data is a distinct artifact, it can be protected, audited, and governed separately from raw data. This is important for regulatory compliance and for enabling AI/ML across administrative domains. The definition also underpins federated learning, where model data (in the form of intermediate models) is exchanged rather than raw training data.

What changed in its specifications

The latest release-to-release comparison of specifications that define AI/ML model data. Each summary describes the whole specification, not only this term.

TS 26.114Rel-18 → Rel-19

Rel-19 adds a reference to the IVAS fixed-point C code specification in TS 26.114. It also lists that specification among supported specifications for MTSI immersive audio clients.

Full comparison in TS 26.114 →

Release Timeline

Evolution Across Releases

An overview written from the defining specifications, not a verified change history. For the exact text, open the specifications listed below.

Rel-19 Initial

Introduced the formal definition of AI/ML model data as the internal components of a trained neural network, including architecture, weights, biases, and parameters, establishing a standardized basis for model management and exchange.

The definition text carried over unchanged from Release 19; model data remains defined as the internal components of a trained neural network that define how inputs are processed to generate outputs.

Defining Specifications

3GPP specifications that define or reference AI/ML model data, with the latest known release. Sourced from the 3GPP document catalog — see methodology.

SpecificationTitleRelease
TS 26.114 vk11 IMS Multimedia Telephony Client Rel-20 read
TR 26.927 vj00 AI/ML in 5G Media Services Study Rel-19 read