GLM-5.3: Zhipu AI's Coding Jumps 50%, Open-Source Soon

Zhipu AI has released its new large language model, GLM-5.3, with a 50% improvement in core coding capability over its predecessor. The company says this leap comes from advanced post-training, not changes to the base model. What's interesting is GLM-5.3 now ranks first among open-source models on several benchmarks, including Terminal Bench 3.0 and Agents' Last Exam. Its coding and agent capabilities are approaching Claude Fable 5. On Terminal-Bench 3.0, GLM-5.3's score jumped from 4.6 to 28.3 points. It also saw significant gains in cybersecurity tests, scoring 84.5% in CyberGym. The model completed 130 vulnerability exploitation tasks in six hours on ExploitGym, a big jump from the previous version. The bottom line: Zhipu AI plans to release the open-source weights for GLM-5.3 in two weeks, making these advanced capabilities more widely accessible.