GLM-5.3: post-training alone 6x'd Terminal-Bench and lifted cyber scores
Quote149 words25 likes
GLM-5.3 keeps the 743B base of GLM-5.2 but gains entirely from post-training: Terminal-Bench 3.0 from 4.6 to 28.3 and DeepSWE from 46.2 to 66.9. Highlights its offensive security results, 84.5% on CyberGym and 2,400+ real vulns disclosed.
Key claims
Terminal-Bench 3.0 4.6 to 28.3, DeepSWE v1.1 46.2 to 66.9
CyberGym 84.5%, ExploitBench 24.4% to 54.4%
2,400+ real vulnerabilities disclosed; weights due in about 2 weeks
GLM-5.3 by @Zai_org is out.
Same 743B base as 5.2 with zero pretraining changes, yet it just 6x'd its score on Terminal-Bench.
All of the gains came from post-training. More environments and longer RL runs.
- Terminal-Bench 3.0: 4.6 to 28.3
- DeepSWE v1.1: 46.2 to 66.9
Now beats Opus 4.8 (while Opus 5 feels like the retarded cousin of the former to me anyhow lol) on their code bench while using fewer output tokens. Fable 5 still leads tho.
But coding gains aside, what actually stands out is the offensive security performance:
- 84.5% on CyberGym. That's above Mythos 5 and GPT-5.6 Sol.
- ExploitBench doubled, 24.4% to 54.4%
- 2,400+ real vulns disclosed across open-source projects
API is live, and weights should drop in 2 weeks.
China AI continues to close the gap, and GLM-5.3 is another strong piece of evidence to this.
Linking this here too:
QUOTING @Zai_org
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models Tech Blog: t.co/ekQkO83jCv t.
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models Tech Blog: t.co/ekQkO83jCv t.