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

Glossary term · Management

AI/ML intermediate model

Artificial Intelligence / Machine Learning Intermediate Model

Management →

AI/ML intermediate model is a partially trained AI/ML model produced during federated learning rounds, not yet satisfying the required training rounds or federation training requirements, and exchanged between federation members and the federation server.

Introduced
Rel-19
Where
Services
Specifications
2 specs
AI/ML intermediate model Description Purpose Neighbours Specifications

Description

The AI/ML intermediate model is a concept introduced in 3GPP Release 19 within the framework of AI/ML management, specifically to support federated learning across the 5G management architecture. In federated learning, multiple federation members (for example, network elements or management functions such as the NWDAF, OAM entities, or RAN nodes) collaboratively train a shared model without exchanging raw training data. Each member trains locally on its own data and periodically sends model updates to a central federation server, which aggregates them into a global model. The intermediate model is the artifact produced at each of these local training steps before the model has converged or before the required number of training rounds has been completed.

Architecturally, the intermediate model sits between the local training process and the federated aggregation function. When a federation member performs a training round, it produces an updated model — the intermediate model — which is then transmitted to the federation server (or to peer members, depending on the federation topology). The server aggregates intermediate models from multiple members to produce a new global model, which is then distributed back to members for the next round. The intermediate model therefore represents a snapshot of the model state at a given point in the training lifecycle, and it is explicitly distinguished from the final or converged model that meets all federation training requirements.

The definition in 3GPP TS 23.482 and TS 24.560 characterizes an intermediate model as one that either has not met the required training rounds or has not yet satisfied the requirements of the federation training process. This distinction matters operationally: management functions need to know whether a received model is an intermediate artifact suitable for further aggregation or a finalized model ready for deployment in inference. The intermediate model carries the model data (architecture, weights, biases, and parameters) along with metadata describing the training round, the member identity, and the training status.

In the broader AI/ML management architecture, the intermediate model enables the federated learning workflow to be managed, monitored, and audited. The management service producer exposes capabilities to transfer intermediate models, and the management service consumer (the federation member or server) uses these to orchestrate training rounds. Because intermediate models may be large, the framework also addresses aspects such as model compression, transfer scheduling, and integrity protection. The concept supports both horizontal and vertical federation scenarios and is designed to be agnostic to the specific ML algorithm used.

The intermediate model is closely tied to the lifecycle management of AI/ML models in 5G and beyond. It allows operators to distribute training across domains (for example, across RAN, core, and OAM domains) while respecting data privacy and locality constraints. By formalizing the intermediate model as a distinct managed entity, 3GPP ensures interoperability between multi-vendor federation members and provides a standardized basis for performance monitoring, fault management, and accounting of federated training procedures.

Purpose & Motivation

Federated learning in telecommunications networks requires multiple participants to train a shared model collaboratively without centralizing sensitive data. Before Release 19, 3GPP management specifications defined AI/ML model management primarily for centralized training and inference, without a standardized notion of the partially trained models exchanged during federated training rounds. This created ambiguity: a management function receiving a model could not determine whether it was a finalized model ready for deployment or an intermediate artifact intended for further aggregation.

The AI/ML intermediate model concept was created to resolve this ambiguity and to enable standardized orchestration of federated learning. By defining intermediate models as those that have not met the required training rounds or federation training requirements, 3GPP provides a clear semantic distinction that supports correct handling, aggregation, and lifecycle management. It also enables multi-vendor interoperability, since federation members from different vendors can exchange and aggregate intermediate models using a common understanding of their status.

Additionally, the concept supports privacy-preserving and distributed AI/ML training, which is increasingly important as operators seek to leverage data across network domains without violating data sovereignty or regulatory constraints. The intermediate model provides the building block for managing training rounds, monitoring convergence, and controlling when a model transitions from intermediate to finalized status.

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 AI/ML intermediate model concept within the AI/ML management framework, defining it as a model that does not meet the required training rounds and/or the requirements of federation training, enabling standardized federated learning orchestration.

The definition text carried over unchanged from Release 19; the term remains defined as a model that does not meet the required training rounds and/or the requirements of federation training.

Defining Specifications

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

SpecificationTitleRelease
TS 23.482 vk30 AI/ML Enablement for Verticals Rel-20 read
TS 24.560 vk10 AIMLE SEAL Services Stage 3 Rel-20 read