In 5G networks,optimization of antenna beam weights of base stations has become the key application of AI for network optimization.For 6G,higher frequency bands and much denser cells are ex-pected,and the importance of automatic and accurate beamforming assisted by AI will become more promi-nent.In existing network,servers are“patched”to network equipment to act as a centralized brain for model training and inference leading to high trans-mission overhead,large inference latency and poten-tial risks of data security.Decentralized architectures have been proposed to achieve flexible parameter con-figuration and fast local response,but it is inefficient in collecting and sharing global information among base stations.In this paper,we propose a novel solution based on a collaborative cloud edge architecture for multi-cell joint beamforming optimization.We ana-lyze the performance and costs of the proposed solu-tion with two other architectural solutions by simula-tion.Compared with the centralized solution,our so-lution improves prediction accuracy by 24.66％,and reduces storage cost by 83.82％.Compared with the decentralized solution,our solution improves predic-tion accuracy by 68.26％,and improves coverage per-formance by 0.4dB.At last,the future research work is prospected.