Zhipu released GLM-5.3 today, with coding and agent capabilities approaching Claude Fable 5, and it has become the strongest open-source security model. Before release, red-team testing with Tsinghua, Nankai, and others found 2,404 vulnerabilities, the oldest dating back 40 years.
SynthePulse Insight · AI deep reading
Version 3 · 4 sources
Zhipu AI releases GLM-5.3, boosting coding ability by 50% through post-training alone, and discovering 2,436 vulnerabilities in cybersecurity. But benchmarks are unverified, ecosystem compatibility remains questionable, and model shelf life shrinks to 20 days.
The most striking aspect of GLM-5.3's release is that it doesn't change the base model but achieves a 50% improvement in coding ability through extreme post-training scaling. Zhipu's official technical blog states that the primary reason for this upgrade is post-training scaling, and notes that efficiently advancing reinforcement learning on the exact same base as GLM-5.2 may be far from reaching the upper bound of that base model's intelligence.
In terms of parameters, GLM-5.3 is comparable to GLM-5.2, at around 700 billion, while contemporaries like Kimi K3 and Qwen3.8-Max have moved toward trillion-parameter models. APPSO's report emphasizes that GLM-5.3 achieves intelligence comparable to Kimi K3's 2.8T-level models with less than half the parameters, which is seen as a victory for post-training scaling.
However, these improvements come from Zhipu's own evaluations or internal benchmarks, not independently verified. Geek Park's report also merely relays Zhipu's claims without third-party test results. Therefore, the 50% improvement and benchmark rankings should be viewed as Zhipu's unilateral assertions.
Another focus of GLM-5.3 is cybersecurity. Zhipu, in collaboration with Tsinghua and Nankai universities, conducted red-team testing, discovering a total of 2,404 vulnerabilities (after initial screening and deduplication), of which 1,088 were medium to high severity, covering 220 projects. The Decoder's report, however, states that GLM-5.3 helped security teams find 2,436 vulnerabilities across 269 projects, with the oldest dating back about 40 years.
In ExploitGym testing, GLM-5.3 completed 130 out of 898 tasks within 6 hours, matching Claude Mythos 5's performance. Qbit's hands-on test showed GLM-5.3 identifying all 12 vulnerability types in the AegisDesk project and fixing them itself, with all 54 regression tests passing in the end.
Notably, the specific methods and verification processes for vulnerability discovery have not been fully disclosed; for example, the discrepancy between 2,404 and 2,436 may stem from different statistical criteria. Additionally, Qbit's test is media testing, not an independent third-party evaluation, so its conclusions should be treated with caution.
APPSO's hands-on testing found that when using GLM-5.3 in Claude Code with automatic command approval enabled, users frequently encounter the error message 'auto mode cannot determine the safety of Bash right now.' This may be a mechanism inherent to Claude Code, but third-party models have not yet adapted.
In contrast, in Zhipu's own ZCode, GLM-5.3 can use tools, take screenshots, and continuously optimize like Codex, but it takes longer. For example, a simulation of a planet hitting Earth took less than 1 hour in Claude Code, but in ZCode it was still running after 1 hour.
Ecosystem compatibility issues may affect user experience, but Zhipu has launched Zcode and AutoClaw, available to GLM Coding Plan users. Third-party platforms like WorkBuddy, QwenWork, and TraeWork have also opened early access, but the API won't be available until next Tuesday.
Zhipu announced that GLM-5.3's weights will be open-sourced within two weeks, along with the post-training framework Slime, which supports GLM, Qwen, some DeepSeek models, and Llama 3. Slime has been used by Zhipu for post-training since GLM 4.5, and its open-sourcing is expected to lower the barrier for post-training.
However, model competition has intensified. APPSO notes that just 20 days after GLM-5.3's release, a stronger Kimi K3 appeared, and 20 days later, Kimi was surpassed by DeepSeek, compressing the model 'shelf life' to about 20 days. This rapid iteration poses challenges for both users and developers.
Despite GLM-5.3's standout performance in coding and cybersecurity, it remains uncertain how long its advantages will last against competitors like Kimi K3, DeepSeek V4, and Grok 4.6. Zhipu's API pricing has not yet been updated, still showing GLM-5.2, but significant changes are not expected.
Qbit's hands-on test showcased GLM-5.3's capabilities on complex tasks: completing the Lord of the Rings benchmark in 28 minutes, delivering a simulated game with backend, database, and permissions in 40 minutes, and proactively distinguishing between VERIFIED, INFERRED, and UNVERIFIED in a GPU-less environment. These performances reflect progress in engineering delivery and risk awareness.
But APPSO's testing showed inconsistent results in 3D generation tasks. For example, a human blood circulation simulation was rendered as a crude 'stick figure' with internal organs piled together, while a jellyfish lake scene was stunning but required a very long prompt. This indicates that the model's performance is highly dependent on prompt quality.
Overall, GLM-5.3 shows significant improvements in coding and security, but real-world effectiveness is heavily influenced by tools, prompts, and scenarios. Users should evaluate its suitability based on specific needs rather than blindly trusting benchmark tests.
This report is based on Zhipu's official statements, hands-on tests by Qbit and APPSO, and reports from The Decoder and Geek Park. All performance data (such as the 50% improvement and vulnerability counts) come from Zhipu or media tests, not independently verified by third parties, and should be considered source claims. Some numbers (e.g., 2,404 vs. 2,436) differ, possibly due to different statistical criteria.
GLM-5.3 demonstrates the potential of post-training scaling, reaching new heights in coding and security, but benchmarks are unverified and ecosystem compatibility needs improvement. In a competitive environment where model shelf life has shrunk to 20 days, its long-term advantages remain uncertain.
Primary report
Primary source
Also covered by 5 sources