Chinese technology startup Z.ai has officially announced the release of its new language model, GLM-5.3. This update marks a significant step forward in programming, agentic scenarios, and vulnerability discovery, strengthening the company's position in the AI solutions market.
Evolution Without Architecture Change
The key feature of the release is an impressive performance gain achieved solely through post-training, without altering the underlying architecture. Compared to the previous version, GLM-5.2, the new model demonstrates improved results on the internal Code Bench benchmark, showing a 50% increase. This points to the high potential of optimizing training algorithms.
Impressive Test Results
Z.ai's release table presents compelling figures across several key benchmarks:
- Terminal Bench 3.0 — 28.3 points versus 4.6 for its predecessor;
- DeepSWE v1.1 — 66.9 versus 46.2;
- Agents' Last Exam — 28.5 versus 23.8;
- CyberGym — 84.5% versus 77.2%;
- ExploitBench — 54.4% versus 24.4%.
ExploitGym deserves special attention: the model completed 105 tasks within a two-hour budget, while GLM-5.2 managed only 29. When the budget was increased to six hours, the figure rose to 130 tasks versus 39 for the previous version. This demonstrates significant progress in autonomous vulnerability discovery.
Practical Application and Security
Z.ai conducted extensive testing on real codebases in collaboration with several security teams in China. After verification and deduplication, the system identified 2,436 vulnerabilities across 269 projects, including 1,097 issues of medium and high severity. Fifty-three findings were publicly disclosed, while the remaining 2,383 remain under embargo. The average "lifespan" of such vulnerabilities is estimated at 26.6 years, with the oldest detected dating back to 1981.
For process transparency, the company launched the Z.ai Security Disclosure Ledger — a public registry tracking the status and severity of findings. In parallel, GLM-5.3 has already been deployed for GLM Coding Plan subscribers, becoming the primary engine for agentic programming and "long-horizon" tasks. Open weights will be published two weeks after the security assessment.
My analytical take: Z.ai demonstrates impressive momentum, especially in cybersecurity, where GLM-5.3's results look revolutionary. However, it is worth noting that the company has previously faced accusations of unauthorized technology distillation, which could affect market trust. Nevertheless, if the model confirms its claimed characteristics in independent tests, this could pose a serious challenge to Western competitors.