Z.ai Launches Enhanced GLM-5.3 Language Model

Z.ai, a Chinese startup, has launched its latest language model, GLM-5.3, which brings significant improvements in programming capabilities, agent scenarios, and vulnerability detection.

The enhancements in GLM-5.3, compared to its predecessor GLM-5.2, were accomplished exclusively through post-training methods without modifying the base model.

The model's open weights will be released two weeks following a security evaluation and subsequent hardening measures.

In internal testing, dubbed Code Bench, GLM-5.3 demonstrated a 50% improvement in coding tasks over GLM-5.2.

Z.ai highlighted several comparative results in its announcement:

• Terminal Bench 3.0: 28.3 points for GLM-5.3 vs. 4.6 for GLM-5.2;
• DeepSWE v1.1: 66.9 vs. 46.2;
• Agents' Last Exam: 28.5 vs. 23.8;
• CyberGym: 84.5% compared to 77.2%;
• ExploitBench: 54.4% vs. 24.4%.

In the ExploitGym assessment, GLM-5.3 successfully completed 105 tasks within two hours, an increase from 29 tasks achieved by GLM-5.2. The model further excelled in a six-hour period, completing 130 tasks compared to 39 previously.

Moreover, Z.ai collaborated with various security teams in China to test the model on actual codebases. This effort led to the identification of 2,436 vulnerabilities across 269 projects, including 1,097 categorized as medium to high severity.

Of these vulnerabilities, 53 have been disclosed publicly, while 2,383 remain confidential. The average age of these vulnerabilities is estimated at 26.6 years, with the oldest dating back to 1981.

In conjunction with this release, Z.ai has introduced the Z.ai Security Disclosure Ledger, a public registry to track the status, affected projects, and severity ratings of the vulnerabilities.

GLM-5.3 is now available to subscribers of the GLM Coding Plan and is being promoted as the main engine for programming agents and handling complex tasks. Requests for older models will automatically be redirected to this new version.

It is important to note that in July, Anthropic accused Z.ai of unauthorized use of GLM-5.2, based on similarities in responses from Claude and GPT.