Zhipu releases GLM-5

Zhipu releases GLM-5.3: without changing the base, only post-training, the throne of open source programming has changed hands. Zhipu officially released GLM-5.3 today. The biggest difference from the previous generation is that the base model has not changed at all, and all improvements come from Scaling in the post-training stage. Through long-range mission environments that are dozens of times larger, richer environment types, and ultra-long post-training time, Zhipu has significantly raised the upper bound of intelligence of the model on the same basis. The company's internal evaluation shows that the new model's programming experience is about 50% better than that of GLM-5.2, making it the most powerful open source model in terms of programming capabilities. On public benchmarks, GLM-5.3 also delivered outstanding results. Terminal-Bench 3.0 score jumped from 4.6 to 28.3, DeepSWE v1.1 rose from 46.2 to 66.9, Agents' Last Exam rose from 23.8 to 28.5, and GDPval-AA v2 reached 1769 points. Its programming and intelligence capabilities are close to Claude Fable 5, and its somatosensory capabilities exceed those of other domestic models. It ranked first in open source in two tests: Terminal Bench 3.0 and Agents' Last Exam (CLI). Post-training Scaling: The potential of the base is far from reaching its peak. Zhipu emphasizes that all the above improvements come from post-training rather than changing models. Based on IndexShare, SAO and the continuously evolving new generation Slime framework, the team efficiently promotes reinforcement learning on the exact same base as GLM-5.2, and admitted frankly that "the intelligent upper bound of this base may be far from developed." This idea sends a clear signal: Competition for large models does not necessarily always rely on stacking parameters. "Practice" the existing base can also approach the cutting-edge level. In the security dimension, GLM-5.3 performs on par with Mythos 5 in tasks such as white-box code review and vulnerability discovery, showing its potential for network defense scenarios. Zhipu will open the complete model weights two weeks after its release, but only after completing the security assessment and model reinforcement to limit potential attack capabilities and retain defensive value. At the same time, the new model will be launched on the official programming tools ZCode, AutoClaw and GLM Coding Plan immediately, and will be available for early access to coding platforms such as Trae, Button, WorkBuddy/CodeBuddy, and Qoder. As open source models approach closed source flagships on the programming track, the second half of the competition between domestic large models is shifting from "who can speak better" to "who can work better." via AI News (au