Just now, GLM-5.3 is here! After acquiring multiple open source SOTAs, I used them to "magically modify" DeepSeek Harness

Zhishi Author | Chen Junda Editor | Xinyuan Zhixi reported on August 14 that just now, Zhipu released its latest generation flagship model GLM-5.3 and announced that the model will be open source within two weeks. This is a model with 743 billion parameters, and like the previous GLM model, it focuses on programming. In multiple mainstream benchmark tests, GLM-5.3 ranks first among all open source or upcoming open source models; its capabilities in programming and agent tasks are close to top overseas models such as Claude Fable 5 and GPT-5.6 Sol. Compared with domestic models, GLM-5.3 and Kimi K3 compete with each other in programming and agent fields, and lead DeepSeek-V4-Pro-0813 in most benchmark tests. In terms of network security capabilities, GLM-5.3 shows certain advantages and is ahead of the other two domestic models. At the same time, GLM-5.3 also achieves a better balance between capabilities and token efficiency. Under the same high-end thinking budget, the accuracy of GLM-5.3 reaches 31.5%, exceeding the 29.5% of Claude Opus 4.8. The average output of the task is about 50,000 tokens, which is less than half of the average output of Opus 4.8. Zhidongxi obtained the internal testing qualification of GLM-5.3 before its release, and was able to experience the complete capabilities of this model in real development scenarios in advance. In our actual testing, the combination of GLM-5.3+ZCode can not only develop a 3D open world driving game from 0 to 1 that replicates the area from Lujiazui to the Bund in Shanghai with real map data at a 1:1 ratio, but can also "nakedly test" the newly open source Harness framework of DeepSeek: understand this code warehouse of over 7,000 files under zero prior conditions, and then create a "personality plug-in" according to its rules. In addition to programming capabilities, Zhipu said that after sufficient pre-training, GLM-5.3 also emerged with good network security capabilities. In cooperation with domestic universities and enterprises, GLM