3GPP TSG-RAN WG2 Meeting #130       	                      R2-2504379
Resubmission of R2-2502732
Malta, MT, 19th-23rd May 2025

Agenda item:	8.1.5
Source:	CMCC, China Telecom, CATT, ZTE, Apple, Samsung, Xiaomi
Title:	Discussion on AI/ML model transfer/delivery
WID/SID:	NR_AIML_air-Core
Document for:	Discussion

Conclusion
Here are the observations and proposals for requirements on AI/ML model transfer/delivery.
Discussion on one-sided model
Proposal 1: There is no requirements on controllability or visibility for UE-sided model transfer/delivery case y, since the model trained in OTT server can be sent to the UE directly from OTT-server (transparent to 3GPP).
Observation 1: For case z1, the model is in proprietary format which is not comprehend by network.
Observation 2: RAN2 focuses on functionality-based LCM during R19 discussion, the network doesn’t be aware of which model is used at UE side.
observation 3: For the visibility and controllability for UE-sided model transfer/delivery case z1, since the model is stored at network side, 
- There is requirement that the network needs to be aware that this is a model (model content or meta info is not aware) since the model is stored at network side 
- There is no strong requirement on visibility and controllability due to its transfer in proprietary format. Whether and when to transfer the model to the UE can be up to network, including network initiates autonomously or based on UE request
Proposal 2: It is proposed to de-prioritize case z1 for one-sided model transfer/delivery, since there is no strong requirement on visibility and controllability for UE-sided model transfer/delivery from the perspective of operator.
Proposal 3: It is proposed to follow case y for UE-sided model transfer/delivery in R19, i.e. the UE-sided model is delivered over the top, which is transparent to 3GPP.

4	

TDoc file reading error

3GPP TSG RAN WG2 #130		R2-2504635
Malta, May 19th – 23rd, 2025

Agenda Item:	8.1.5
Source:	BT Plc, Turkcell, T-Mobile, Deutsche Telekom, Orange
Title:	Operators views for the AI_ML model delivery options
Document for:	Discussion/Decision

Conclusions
During the TDoc, we have made the following observations

Observation 1: There are no standardized test to validate AI/ML models
Observation 2: AI/ML models are expected to be created by infraestructe vendors, device vendors and chipset vendors
Observation 2: AI/ML models may require to be deactivated if the end-to-end performance is impacted

For those reasons, it is proposed

Proposal 1: Solution 4a, OTT without 3GPP impact, captured on [1] is no further discussed in RAN2
Proposal 2: One-sided dataset/model transfer solution shall follow below principles:
	A1 - Size: from RAN2 point of view, aim to support various sizes of dataset/model parameter transfer (FFS on model size);
	A2 - Continuity: service continuity of model transfer/delivery during UE mobility needs to be supported;
	A3 - Controllability: NW decides on if and when to transfer/delivery over the air interface;
	A4 - Latency: relaxed latency requirement and infrequent update;
	A5 - Visibility: models to be understandable by MNO;
Proposal 3: Adopt additional list of requirements for one-sided and two-sided model transfer/delivery.
	Activation: once the AI/ML one-sided or two-sided model is transferred to NW and/or to UE, MNO should be capable to activate and deactivate them at any point in time
	Model transfer/delivery: traffic should be transferred at a different priority, e.g., lower than user traffic.
	Differentiability: model transfer/delivery traffic should be differentiated from other user traffic.
	Security: there should be a guarantee that models are transferred securely, in a NW-aware manner, such that untrusted models cannot be downloaded.
	Addressability: Models need to be addressable such that the UE can request the transfer/delivery of a specific one
	Controllability: the NW is in control of if and when to transfer/deliver a model

3GPP TSG RAN WG2 #130		R2-2504641
Malta, May 19th – 23rd, 2025

Agenda Item:	8.1.5
Source:	BT Plc, Turkcell, T-Mobile, Deutsche Telekom, Orange
Title:	Operators views for the AI_ML model delivery options
Document for:	Discussion/Decision

Conclusions
During the TDoc, we have made the following observations

Observation 1: There are no standardized test to validate AI/ML models
Observation 2: AI/ML models are expected to be created by infraestructe vendors, device vendors and chipset vendors
Observation 2: AI/ML models may require to be deactivated if the end-to-end performance is impacted

For those reasons, it is proposed

Proposal 1: Solution 4a, OTT without 3GPP impact, captured on [1] is no further discussed in RAN2
Proposal 2: One-sided dataset/model transfer solution shall follow below principles:
	A1 - Size: from RAN2 point of view, aim to support various sizes of dataset/model parameter transfer (FFS on model size);
	A2 - Continuity: service continuity of model transfer/delivery during UE mobility needs to be supported;
	A3 - Controllability: NW decides on if and when to transfer/delivery over the air interface;
	A4 - Latency: relaxed latency requirement and infrequent update;
	A5 - Visibility: dataset and model parameters to be visible and understandable by MNO;
Proposal 3: Adopt additional list of requirements for one-sided and two-sided model transfer/delivery.
	Activation: once the AI/ML one-sided or two-sided model is transferred to NW and/or to UE, MNO should be capable to activate and deactivate them at any point in time
	Model transfer/delivery: traffic should be transferred at a different priority, e.g., lower than user traffic.
	Differentiability: model transfer/delivery traffic should be differentiated from other user traffic.
	Security: there should be a guarantee that models are transferred securely, in a NW-aware manner, such that untrusted models cannot be downloaded.
	Addressability: Models need to be addressable such that the UE can request the transfer/delivery of a specific one
	Controllability: the NW is in control of if and when to transfer/deliver a model