GLM-5.3: Z.ai bets on post-training, but the weights wait on safety

Z.ai has announced GLM-5.3, an update that leaves GLM-5.2's base model untouched and bets everything on post-training to push performance higher. The company claims substantial coding gains, with a 50% improvement on its internal Z.ai Code Bench and higher scores on tests such as Terminal-Bench and DeepSWE. With no independent evaluations available at launch, though, these figures remain self-reported and should be treated with due caution. Scaling the post-training relied on the open source RL framework 'slime', paired with Megatron for training and SGLang for rollout.

The most striking part concerns cyber capabilities. Working with unnamed Chinese security teams on real codebases, the model reportedly identified — after expert review, screening and deduplication — 2,436 vulnerabilities across 269 open source projects, 1,097 of them of medium-to-high severity. On CyberGym it claims 84.5% against GPT-5.6 Sol's 83.6% (77.2% for GLM-5.2); on ExploitBench it reaches 54.4%, more than double the previous version's 24.4%, yet still far from GPT-5.6 Sol's 76.5%. Of that huge pile of flaws, only 53 are currently visible in the public registry, while the remaining 2,383 sit under embargo and the teams involved stay anonymous, which limits outside verification.

For the first time, Z.ai has decided to postpone the weight release, now expected in about two weeks, in order to complete its “safety evaluation” procedures. The company justifies this departure from its usual policy of immediate openness by pointing to the model's potential dual use. Yet it remains an account the company gives of itself: whether offensive capability really grew beyond expectations cannot be checked from outside, because the only evidence is the story Z.ai tells about its own post-training.

The technical rollout imposes new constraints on developers: the API no longer supports turning reasoning off, making `enabled` mode mandatory even at reduced effort. On top of that, the GLM Coding Plan introduces a points-based quota that discourages peak-hour use, with calls outside the 14:00–18:00 (UTC+8) Monday-to-Friday window — whole weekends included — charged at half the points.

— Olya There is a certain elegance to the delay: to sell us a model that can find flaws nobody spotted in decades, Z.ai first has to make sure it hasn't accidentally handed over the weapon to exploit them. Whether the weights arrive in two weeks or two months, the real news is that total openness has run into a limit called responsibility.