Description
The Model Training Logical Function (MTLF) is a key architectural component introduced in 3GPP Release 17 as part of the 5G system's support for network automation and data analytics. It is defined within the Management Data Analytics (MDA) framework and operates as a logical function that can be deployed within the 5G Core Network or in a management system. The primary role of the MTLF is to ingest, process, and analyze network data to train, validate, and produce machine learning (ML) models. These models are then used to optimize network performance, predict failures, manage resources, or enhance user experience. The MTLF interacts with other network functions and data sources through standardized interfaces to collect training datasets, which may include performance measurements, configuration data, fault information, and user plane analytics.
The MTLF operates through a defined workflow that includes data collection, model training, and model publication. It receives data from various producers, such as the Network Data Analytics Function (NWDAF), Operations, Administration and Maintenance (OAM) systems, or other network functions (NFs) via the Nnwdaf_MTLtraining_Request service operation. The data is formatted according to analytics subscriptions and can be raw or pre-processed. The MTLF then applies machine learning algorithms—which are implementation-specific but could include regression, classification, clustering, or deep learning techniques—to this data to create a trained model. The training process may involve feature extraction, model selection, hyperparameter tuning, and validation against test datasets to ensure accuracy and avoid overfitting.
Once a model is trained and meets required performance metrics, the MTLF publishes it to a Model Repository Function (MRF) or directly to a consumer, such as an NWDAF instance that will perform inference. The model is typically represented in a standardized format like the Predictive Model Markup Language (PMML) or the Open Neural Network Exchange (ONNX). The MTLF also manages the lifecycle of these models, including versioning, retraining triggers (e.g., based on data drift, periodic schedules, or performance degradation), and retirement. It can be configured with training policies that define objectives, data requirements, and performance thresholds.
Architecturally, the MTLF is part of the broader data-driven ecosystem in 5G. It works in concert with the NWDAF (which focuses on analytics inference and exposure), the MRF (for model storage), and the OAM system. The interfaces like Nnwdaf_MTLtraining_Request/Response (defined in TS 29.520) facilitate communication between an NWDAF (as a consumer) and the MTLF. This separation of training (MTLF) and inference (NWDAF) allows for scalable, specialized deployments where computationally intensive training can be offloaded to dedicated platforms, while lightweight inference occurs closer to the network edge. The MTLF enables use cases such as predictive load balancing, anomaly detection, energy saving, slice-specific optimization, and Quality of Experience (QoE) prediction.
Purpose & Motivation
The Model Training Logical Function was created to address the growing complexity of 5G networks and the need for intelligent, automated management. Traditional network management relied on manual configuration and rule-based automation, which could not efficiently adapt to dynamic conditions, predict issues, or optimize performance in real-time. The explosion of data from network functions, devices, and services presented an opportunity to leverage machine learning, but there was no standardized way to integrate ML model training into the network architecture.
MTLF was motivated by the vision of self-organizing networks (SON) evolving into AI-native networks. It solves the problem of how to systematically generate and update ML models using live network data within a standardized framework. Before MTLF, ML capabilities were vendor-specific, proprietary solutions that lacked interoperability and made it difficult to ensure consistent model quality or share models across different network domains. The MTLF provides a standardized, open interface for requesting model training, enabling network functions like NWDAF to consume analytics models without being tied to a specific vendor's training platform.
Its introduction in Release 17, as part of the 5G Phase 2 enhancements, specifically supports advanced network automation scenarios defined in the 5G System Architecture (TS 23.501) and the Management and Orchestration (MANO) framework. It allows operators to deploy closed-loop automation where analytics insights from trained models directly drive network actions (e.g., via the OAM or policy control). This is critical for managing network slicing, where each slice may require unique performance models, and for meeting the stringent latency, reliability, and efficiency demands of 5G verticals.
What changed in its specifications
The latest release-to-release comparison of specifications that define MTLF. Each summary describes the whole specification, not only this term.
Rel-19 added MPQUIC steering to multi-access PDU sessions, XR and media handling, and an Energy Information Function. It also added Local Offloading Management for edge computing, Vertical Federated Learning for AI/ML, and Ambient IoT text.
Full comparison in TS 23.501 →Rel-19 adds Vertical Federated Learning training and inference APIs to TS 29.520, along with signalling storm, QoS/policy, and roaming analytics subscriptions. It also extends NWDAF ML model provisioning, event subscription fields, and 308 Permanent Redirect responses.
Full comparison in TS 29.520 →Technology family
Network exposure and analytics · Core Network
- Rel-11 SCS/AS
- Rel-13 IWK-SCEF, SCEF
- Rel-15 CAPIF, NEF, NWDAF
- Rel-16 SCEF+NEF
- Rel-17 ADRF, AnLF, DCCF, L-NEF, MDAF, MFAF, MTLF
- Rel-18 RE-NWDAF
- MTLF part of NWDAF
“When one or more trained ML Models are available for one or more Analytics ID(s) the NWDAF containing MTLF shall include the Analytics ID(s) that is(are) supported per service in the registration towards NRF.” TS 23.288
Members and relations were proposed independently by several models and kept only where at least two agree; quotes are verified verbatim against the specifications. Release = first approved version of the defining specification.
Release Timeline
Detected Changes Across Releases
verified against 3GPP sourcesDocumented changes in 4 releases, each verified against the text of 3GPP specifications or approved change requests and shown with the exact wording. The release timeline above shows the same releases. A release without an entry has no documented change for this term.
MTLF is defined for the first time as a logical function in NWDAF that trains ML models and exposes ML model provisioning and info services, with registration, discovery and retraining procedures.
- “Model Training logical function (MTLF): A logical function in NWDAF, which trains Machine Learning (ML) models and exposes new training services” TS 23.288 v17.12.0
- “an NWDAF containing MTLF shall include the ML model provisioning services (i.e. Nnwdaf_MLModelProvision, Nnwdaf_MLModelInfo) as one of the supported services” TS 23.288 v17.12.0
- “is provided by the Network Data Analytics Function (NWDAF) containing Model Training Logical Function (MTLF).” TS 29.520 v17.19.0
In Rel-18 the NWDAF containing MTLF gained federated learning server/client roles, discovery and registration with Analytics IDs and interoperability indicators, accuracy monitoring support, and new training service operations.
- “If the NWDAF containing MTLF can act as an FL server for the ML Model training, then FL procedure is initiated by the NWDAF containing MTLF” TS 23.288 v18.13.0
- “The ML Model Interoperability indicator comprises a list of NWDAF providers (vendors) that are allowed to retrieve ML Models from this NWDAF containing MTLF.” TS 23.288 v18.13.0
- “The Network Data Analytics Function (NWDAF) containing MTLF supports (un)subscription to the notification of analytics accuracy monitoring of an ML Model and analytics feedback information.” TS 29.520 v18.15.0
NWDAF containing MTLF gains new Rel-19 functions: training ML models for LMF-based AI/ML positioning, monitoring ML model performance for the LMF, and supporting VFL with an AF as VFL server.
- “For ML model training for LMF-based AI/ML positioning, the NWDAF containing MTLF collects input data from the LMF as described in TS 23.273 [39].” TS 23.288 v19.8.0
- “The NWDAF containing MTLF provides a new or retrained ML model to the LMF, or notifies the LMF that the ML model is degraded” TS 23.288 v19.8.0
- “The NWDAF containing MTLF decides to use VFL for the subscription and realizes it cannot be VFL server but wants an AF to act as VFL server” TS 23.288 v19.8.0
MTLF ML model accuracy monitoring was extended to cover loss values and loss thresholds in addition to accuracy, and VFL-related training handling by the NWDAF containing MTLF was revised.
- “which includes accuracy value or loss value and its compute method of the ML Model and optionally the corresponding number of samples” TS 23.288 v20.2.0
- “the NWDAF containing MTLF reports the ML Model accuracy or ML Model loss to NWDAF containing AnLF either periodically” TS 23.288 v20.2.0
- “indicating the accuracy threshold or loss threshold of the ML Model requested by the consumer” TS 23.288 v20.2.0
Defining Specifications
3GPP specifications that define or reference MTLF, with the latest known release. Sourced from the 3GPP document catalog — see methodology.
| Specification | Title | Release |
|---|---|---|
| TS 23.501 vk30 | 5G System Architecture | Rel-20 read |
| TS 23.700 vk10 | AI/ML Application Layer Support Phase 2 | Rel-20 read |
| TS 29.520 vk10 | 5G NWDAF Services | Rel-20 read |
| TS 29.552 vk10 | 5G Network Data Analytics Procedures | Rel-20 read |