Artificial Intelligence / Machine Learning Model

Description

An AI/ML Model, as defined in 3GPP Release 19 and refined in Release 20, is a set of parameters — including AI/ML model data and associated metadata — required by an AI/ML inference engine to execute an AI/ML task. This definition appears across a range of specifications, including TS 26.114 (media handling), TS 26.927 (AI/ML for media), TS 38.300 (NR overall description), and TS 38.843 (AI/ML for NR air interface). The model is the core artifact that encapsulates learned behavior and enables inference in live networks.

Architecturally, an AI/ML Model consists of two main parts: the model data and the metadata. The model data represents the learned parameters — for example, weights and biases of a neural network — that define the mapping from inputs to outputs. The metadata describes how the model should be used: its input and output formats, the task it addresses, performance characteristics, versioning information, and any constraints. This separation allows an inference engine to load and execute the model correctly without hardcoding assumptions about its structure. The inference engine itself is the runtime component that feeds inputs to the model and produces outputs.

In 3GPP networks, AI/ML Models can be deployed in multiple locations. In the RAN, they may reside in the gNB or be split between the UE and the network, supporting use cases such as channel state information (CSI) feedback enhancement, beam management, positioning, and mobility optimization. In the core network, they support network data analytics (NWDAF) and policy decisions. In media services, they enable enhanced codecs, quality adaptation, and immersive experiences. The specifications define how models are trained, distributed, activated, and monitored, including lifecycle management and performance tracking.

The Release 19 definition described an AI/ML Model as "a data driven algorithm that applies AI/ML techniques to generate a set of outputs based on a set of inputs." This early wording emphasized the algorithmic nature of the model. Release 20 changed the wording to focus on the model as a concrete set of parameters and metadata required by an inference engine. This shift reflects a more implementation-oriented view: the model is not just an abstract algorithm but a deployable artifact with defined interfaces, enabling standardized model transfer, execution, and interoperability across vendors.

The evolution toward a parameter-centric definition supports practical concerns such as model storage, versioning, and exchange between training and inference entities. It also aligns with efforts to define model identifiers, performance metrics, and monitoring mechanisms. As AI/ML becomes integral to 5G-Advanced and 6G, the AI/ML Model will be a foundational building block for intelligent, adaptive network behavior.

Purpose & Motivation

The AI/ML Model concept exists to provide a standardized, interoperable representation of machine learning artifacts within telecommunications networks. As 3GPP began incorporating AI/ML into RAN, core, and media services, it became clear that a common definition was needed to describe what a model is, how it is used, and how it can be exchanged between entities. Without this, each specification area might have developed incompatible notions, hindering multi-vendor deployment and lifecycle management.

The term solves several problems. First, it separates the model from the inference engine, allowing the same model to be executed on different platforms. Second, it introduces metadata as a first-class component, enabling automated discovery, validation, and compatibility checks. Third, it supports model lifecycle operations such as training, testing, deployment, activation, and deactivation, which are essential for reliable AI-driven features. These capabilities were absent in earlier, algorithm-centric descriptions.

The Release 20 rewording from "data driven algorithm" to "a set of parameters ... required by an inference engine" marks a maturation of the concept. It reflects lessons learned from early AI/ML studies, where ambiguity about model boundaries and interfaces caused integration challenges. By defining the model as a concrete, parameterized artifact with metadata, 3GPP enables standardized model transfer, performance monitoring, and multi-vendor interoperability, laying the groundwork for AI-native network design.