Chinese AI startup Zhipu, which evolved from Tsinghua University research, released GLM-5.3 on August 14.

The company said the base model is unchanged from GLM-5.2, but additional post-training raised its performance ceiling.

Zhipu also said it plans to open the model weights two weeks after the release.

Zhipu said GLM-5.3 improved programming capabilities by 50% vs. GLM-5.2 in an internally built subjective evaluation.

It also said the model ranked first among open-source models on public benchmarks, including Terminal Bench 3.0 and Agents’ Last Exam (command-line interface, or CLI).

Earlier GLM-5 documentation described Terminal-Bench as measuring real-world terminal tasks.

Zhipu was added to the US Entity List on January 16, 2025, requiring licenses for exports subject to US rules, with a presumption of denial.

🔗 Source: Yicai

Recent Zhipu developments