Originally published on AI Tech Connect.
What you need to know The base model did not change. Z.ai shipped GLM-5.3 on 14 August 2026 on the same base as GLM-5.2, and says the sharp movement in coding and agent scores came from post-training. The benchmark number. On CyberGym, which scores how well a model finds known vulnerabilities in source code, Z.ai reports 84.5 per cent against 77.2 per cent for GLM-5.2. These are vendor-reported results. The disclosure ledger. Z.ai says its models have surfaced 2,436 vulnerabilities across 269 open-source projects since GLM-5.2, of which 1,097 were rated critical or high severity. No third party has audited it. The weights are not out. Z.ai said they would follow in roughly two weeks, after safety evaluation and hardening — the first GLM release held back explicitly for a safety review.…
Top comments (0)