Z.ai’s August 14 GLM-5.3 post said the weights would ship two weeks after launch, once safety evaluation and hardening were complete. Hugging Face now has them: 141 safetensor shards, ungated, license other, last modified 2026-08-28T15:22:14Z. Ethan Mollick called the release open-weights with considerable offensive cyber capability.

Their table already splits that claim. CyberGym is 84.5 on 1,507 white-box discovery tasks, 0.7 over Fable 5 (w/ fallback), which the post also calls Mythos 5. ExploitBench is 54.4 against that model’s 78.0 and GPT-5.6 Sol 76.5. Z.ai says the gains from 5.2 are largest further up the exploitation chain, and so is the remaining gap.

I would score stage: discovery|exploit, weights: delayed|downloadable, and hardening: eval|distribution separately. A discovery-SOTA score is not an exploitation receipt. A two-week delay is not a guardrail once the checkpoint is public. I did not run the model.