Z.ai releases GLM model that gains coding skill and security strength from extra training

Z.ai releases GLM model that gains coding skill and security strength from extra training

An updated version of GLM shows clear gains on complex programming work and unexpected ability to locate weaknesses in software, with public release of its parameters planned after safety checks.

GP
Giulio Prisco
Aug 17, 2026
2 min read

Z.ai has released GLM-5.3, based on the same underlying model as the previous version, GLM-5.2. Reported improvements come only from additional training after the main model was already complete. This later stage of post-training focused on giving the model more practice with varied and extended tasks. The developers describe building larger collections of practice environments and spending more computing power on this stage.

The model now handles complex programming work and tasks that unfold over many steps more effectively than before. On several standard tests that measure coding skill, it records higher scores and is described as the strongest publicly available model of its kind in this area.

Performance gains and cybersecurity findings

During the same extra training, the model developed a stronger ability to examine computer code for security problems. This strength appeared without being the main goal of the training. On tests that measure finding weaknesses in software, the model reached the highest scores among those compared. It also improved on tests that go further and examine how those weaknesses might be used. In practical checks, the model located thousands of real weaknesses across hundreds of existing software projects, some of them decades old.

GLM-5.3 has "top-tier coding and agentic capabilities," the company posted on X. "A major leap in cybersecurity, setting a new standard among open models." In another post, the company stated that "GLM-5.3 takes agentic coding to the next level, delivering a dramatic improvement over GLM-5.2 while achieving better results with fewer output tokens."

The announcement states that the full set of model parameters will be made public about two weeks after the initial release. This will happen once further safety checks and protective measures are finished. Until then the model is available through a paid coding service. The post presents these results as evidence that careful scaling of the later training stage alone can produce useful gains in both everyday programming ability and specialized security analysis.

About the Writer

More from Mindplex

Keep reading

Three more ideas worth your time.

Browse News

Discussion

Join the discussion

Sign in to share a response with the community.

Type @ to mention someone Type / or use + to add a block Highlight text, then choose Link
Loading editor

Comments cannot be edited after posting because they become part of the reputation record. Give yours a quick review first.