Based on the earlier study in phase one TR 22.874, operators can provide services to help manage and distribute the AI/ML models especially in the edge server so that the UE can acquire a proper model immediately. However, when a lot of UEs requesting for the same model at the same time or the UE is blocked by barriers with poor connection with the base station, the model transfer process will become longer than expected. To overcome this difficulty, as shown in Figure 1, a volunteer UE which is well connected to the base station can help relaying AI/ML models or receive and store AI/ML models first. Then, the other UEs can download AI/ML models from the volunteer UE through direct device connection. In this way, all UE can have a stable and reliable model transfer process while the radio resource of the base station can be saved. Besides, the volunteer UE can transfer the stored models to other volunteer UEs under operator's control. The selection of volunteer UE can be realized by local network policies and strategies. And it also can be exposed as a capability to the 3rd party company when the company wants to choose one or a few certain UEs to be volunteer UEs in an activity. For example, a travel company may assign the tour guides' Augmented Reality (AR) headsets as volunteer UEs in a carnival through the operator's network exposure. The travel company may sign a higher quality plan for tourist guides' devices to provide better user experience for following tourists. Meanwhile, operator can benefit from the alternative open service based on AI/ML model management capabilities and may avoid low Quality of Service due to crowding direct connections to base stations during the carnival.

The operator's MEC near the Jurassic Park stores a variety of AI/ML models according to the the park company's requirements. And it is capable to transfer the stored model to the device such as AR headset. The operator rolls out a new high-quality plan which can allow the user customizes own Service Level Agreement (SLA) for specific network address access and data (e.g. AI/ML Models) download. As a trade-off, the user's device will help transfer the same data through direct device connection to nearby devices sharing common aspiration. The AR headset can transfer the stored AI/ML model to the other AR headsets. However, the AR headset cannot store all models for different scenarios due to limited storage. Indeed, a model needs to be downloaded when or a few seconds before the UE first appears in the certain area. Alice and Bob are tour guide hired by Jurassic Park and their real-time positions can be acquired when they are in the park based on signed agreements. All of AR headsets should in the coverage of the base stations serving for the Jurassic Park.

1. Jurassic Park provides panorama AR tour guide services in a commercial area and a tropical rainforest area. AR headsets need to download Model A and B (both are VGG-16, 552MByte) respectively.
2. To provide high quality user experience, Jurassic Park company indicates to the operator that AR headsets require to download model A in area A and model B in area B.
3. The Jurassic Park company signs a high-quality plan for tour guide Alice and Bob's AR Headsets for providing better service to the tour group using direct device connection.
4. When Bob and his tour group enter area A, their headsets request for the Model A. The operator network finds they requested the same model and Bob is a signed volunteer UE, then triggers to establish a QoS acceleration for Bob's model downloading timely within 1 second. Meanwhile, Jurassic Park requests the operator network to inform Bob to help transfer the model to all other UE near Bob. Also, the operator network informs all other UE near Bob that Bob can provide the model as well. The UE which is a little far from Bob (e.g. out of Bob's coverage) will still download the model through the base station directly.
5. Alice and her group are 10 meters far from Bob and also move towards to area A. Jurassic Park predicts their desire model based on their movement and finds Bob has already downloaded it based on the model transferring records. Jurassic Park requests operator network to inform Alice that she can request model from Bob. Meanwhile, the operator network indicates all other UE near Alice to download the model from Alice.
6. For Alice and Bob, they can see the status of all direct device connections to themselves through network exposures providing by operators (e.g. monitored bandwidth and latency of each direct device connection).
7. When Alice and Bob notice that their groups have a poor QoS of model transfer through direct device connection, they can send a request to the park company for promoting the performance of their direct device connections and the park company will send a similar message to the operator through network exposure to active a temporary acceleration of these direct device connections (e.g. expand the bandwidth of each direct device connection).

1. The tourists can enjoy the continuous AR services with smooth model switchover when their location and responding models change.
2. Tour group's AR headsets provides user experience of the panorama AR tour guide services that can help retrain and improve AI/ML models in operator's MEC by Jurassic Park company (e.g. Federated/Distributed Learning).
3. the operator network performs analytics, based on network statistics and quality of experience reported by Jurassic Park company, to improve and optimized the model transfer process (e.g. setting constraints for maximum direct device connection for one volunteer UE and choose a temporary volunteer UE for sharing model transfer task).

In clause 6.27.2 of TS 22.261 "Requirements" In clause 6.9.2.4 of TS 22.261 "Relay UE Selection" In clause 6.40.2 of TS 22.261 v18.6.0