Z.ai holds GLM-5.3 weights for safety review

Z.ai says its new GLM-5.3 lifted its internal coding score 50% over GLM-5.2 and reached 84.5% on CyberGym vulnerability discovery, but the company is holding the weights for roughly two weeks of safety evaluation and hardening. TechNode’s launch report and InfoWorld’s account say the model trails leading systems on deeper exploit tests, making the delay a governance decision rather than a routine open-weights launch. Security and coding teams should separate defensive benchmark claims from deployable weights and add a release gate, alongside the archive’s warning that GPU-code scores can overstate correctness.