Z.ai publishes GLM-5.3 weights and reports large gains on vulnerability discovery benchmarks

Sources: model card, discussion
Summary: The model card states that GLM-5.3 uses the same base model as GLM-5.2 and that every gain comes from post-training, and it publishes a 753 billion parameter FP8 checkpoint across 141 shards under a license named glm-5.3 rather than a standard open-source license. Z.ai labels the largest change an emergent cyber capability that developed faster than expected, and reports CyberGym at 84.5 against 77.2 for GLM-5.2, ExploitBench at 54.4 against 24.4, ExploitGym at 105 solves in two hours and 130 in six against 29 and 39, and Terminal Bench 3.0 at 28.3 against 4.6. Every figure is vendor-reported with no independent reproduction located, and third-party serving is already live through Baseten and Z.ai.
Why it matters: Downloadable weights that the vendor itself positions on vulnerability discovery and exploitation change the threat model for who can run a capable offensive agent.
Follow-up: Track independent reproduction of the CyberGym, ExploitBench and ExploitGym figures, and the terms of the glm-5.3 license.