Large language model developer Z.AI, previously known as Zhipu AI, has announced the release of GLM-5.3, claiming it ranks as the top open-source model across several major benchmarks. The company states that its coding and AI agent functionalities are nearing the sophistication of Claude Fable 5, one of the most advanced models from a notable AI firm.
GLM-5.3 shares the same foundational architecture as the earlier GLM-5.2 but has undergone extensive post-training expansion, significantly boosting its intelligence potential, the Beijing-based company said. This open-source model offers enhanced coding and cybersecurity features and provides an improved user experience compared to other Chinese-developed models.
The high benchmark scores achieved by Z.AI are significant because its open-source model is closing the performance gap with leading proprietary models from prominent firms, particularly in complex coding and AI agent tasks. Unlike proprietary options, open models allow developers more flexibility to customize and deploy independently. Other Chinese open models include offerings from Moonshot AI and DeepSeek.
On the Terminal-Bench 3.0, which assesses a model’s ability to complete sophisticated tasks in a real terminal environment, GLM-5.3 scored 28.3 percent—up from just 4.6 percent for its predecessor. In comparison, Claude Opus 5, a closed-source model from another firm, scored 42.7 percent and ranks highest on this benchmark. Meanwhile, GLM-5.3’s score on DeepSWE v1.1, which measures long-term software engineering and continuous code alteration skills, increased to 66.9 percent from 46.2 percent, approaching the top score of 74 percent achieved recently by Claude Opus 5.
In the Agents’ Last Exam, which tests real-world professional scenarios involving cross-tool collaboration and extended tasks, GLM-5.3 improved to 28.5 from 23.8. Additionally, on the GDPval-AA v2 test that covers 44 occupations and assesses practical, high-value knowledge work, the model scored 1,769 points—placing it roughly 4 percent below Claude Opus 5 Max’s 1,849 points. This indicates GLM-5.3’s growing capacity to handle specialized tasks leveraging its coding expertise, said the company.
The competition among major AI developers has recently intensified, with several new models entering the scene. DeepSeek launched V4 Pro this week, making it accessible for developers via an API. This new version, DeepSeek-V4-Pro-0813, boasts improved AI agents and capabilities.
Earlier this week, SpaceXAI, the AI division of Elon Musk’s space exploration company, introduced Grok 4.6, an upgrade focused on enhancing performance for long-horizon tasks and managing more complex interactive, visual, and knowledge-intensive operations.
Despite these launches, Z.AI’s shares [HK: 2513] closed 3.6 percent lower at HKD 1,270 (about USD 162) today, though they are still nearly 11 times the price at their initial public offering in January.