Crypto news

Chinese startup Z.ai has officially unveiled its GLM-5.3 language model, betting on three key areas: coding, autonomous agent scenarios, and cybersecurity. This is not just an incremental update—developers are claiming a radical performance leap across several specialized benchmarks.

Post-training without architecture changes

Notably, the increase in computational power was achieved solely through post-training, without altering the base model compared to GLM-5.2. This points to significant potential in optimizing training algorithms, which is rare in the industry. Open weights will be available in two weeks, following the completion of safety assessments and additional hardening.

On the internal Code Bench test, which evaluates coding quality, the new model improved its predecessor's result by 50%. However, the data for agent-based and cybersecurity scenarios is far more impressive:

• Terminal Bench 3.0 — 28.3 points versus 4.6 for GLM-5.2 (more than a sixfold increase);
• 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%.

Hunting for vulnerabilities: 2,436 findings

ExploitGym deserves special attention. Within a two-hour budget, the model completed 105 tasks versus 29 for GLM-5.2, and with a six-hour budget, 130 versus 39. This demonstrates not just improvement, but a qualitative shift toward real autonomous operation.

During testing on real codebases in collaboration with Chinese security teams, the system identified 2,436 vulnerabilities across 269 projects, including 1,097 issues of medium and high severity. Only 53 findings were publicly disclosed, with the remaining 2,383 kept under embargo. The average "age" of detected vulnerabilities is 26.6 years, with the oldest dating back to 1981.

For process transparency, Z.ai launched the Security Disclosure Ledger—a public registry tracking the disclosure status and severity of each finding. This is a rare step for the industry, enhancing trust in the model.

Product launch and market context

GLM-5.3 is already deployed for GLM Coding Plan subscribers as the primary engine for agentic coding and long-running tasks. Requests to older models are automatically routed to the new one, easing user migration.

Recall that in July, Anthropic accused Z.ai of unauthorized distillation of GLM-5.2 using responses from Claude and GPT. This release is likely to intensify competitive pressure in the AI agent segment for developers.

My analysis: Growth in agent scenarios and cybersecurity is a strategic move by Z.ai to differentiate from OpenAI and Anthropic, which focus on general-purpose models. However, the 50% improvement on Code Bench, achieved without an architecture change, raises questions about the reproducibility of such results in real-world conditions. The industry needs independent benchmarks, not just internal metrics.