THINK SPECTRUM: RELEASE OF THE GLM-5.3 MODEL PERFORMED BETTER IN COMPLEX CODING AND LONG-RANGE TASKS

On 14 August, we announced that we had the GLM-5.3 today. It uses the same basic model as GLM-5.2 - all performance upgrades are derived from post-training optimization. Compared to GM-5.2, it performed better in terms of complex coding and long-range tasks. GLM-5.3 has a stronger coding capability and is currently the most powerful open-source weight model, which in our internal Z.ai code benchmark test is 50% more effective than GLM-5.2. Moreover, when the training is followed by a scalable deployment, its network capacity has developed faster than we expected. On the CyberGym platform, GLM-5.3 leads in the detection of loopholes, and its most significant increase in the back of the utilization chain - more than twice as high as the GLM-5.2 in the use of baseline tests. We will announce the weights two weeks after the release, after the completion of the security assessment and enhancements. (thinking)