Zhipu releases GLM-5.3, takes top open-source spot in coding and cybersecurity
On August 14, Zhipu released GLM-5.3, a 743B-parameter model whose gains came entirely from post-training scaling. It set the open-source record on Terminal-Bench 3.0 and CyberGym, with coding capability approaching Claude Fable 5.
On August 14, Zhipu AI released GLM-5.3, a 743-billion-parameter base model that shares its foundation with GLM-5.2 — every improvement comes from post-training scaling. Zhipu explicitly stated the base model is unchanged; the intelligence ceiling was raised through more demanding long-horizon environments, richer environment types, and significantly longer post-training.
On coding, GLM-5.3 set a new open-source record of 28.3 on Terminal-Bench 3.0, surpassing Kimi K3, and scored 66.9 on DeepSWE 1.1, putting its real-world coding feel on par with Claude Fable 5 and GPT-5.6 Sol. At the High reasoning setting, GLM-5.3 reached 31.4 percent accuracy versus Claude Opus 4.8's top-tier 29.5 percent, while consuming only about 50,000 tokens per task on average compared to Opus 4.8's roughly 120,000 tokens, indicating shorter execution paths to completion.
The cybersecurity gains drew more attention. On CyberGym, which evaluates vulnerability reasoning from white-box source code by triggering program faults, GLM-5.3 scored 84.5 percent, ahead of GLM-5.2 at 77.2, Mythos 5 at 83.8, and GPT-5.6 Sol at 83.6. Two weeks before launch, Zhipu partnered with domestic security teams to discover 2,404 deduplicated potential vulnerabilities, of which 1,088 were medium-to-high severity. Researchers using GLM-5.3 uncovered a 40-year-old DNS protocol-level risk that can amplify server load by up to nearly 80,000 times with a small number of crafted requests, affecting more than 10 million public DNS services.