GLM-5.3 is out.
To be precise, the announcement is out.
Z.ai's official post says it will release the weights two weeks later, after a safety review.
Miss that difference and the headline "open model released" runs ahead of reality.
Same Base as GLM-5.2, Different Post-Training
Z.ai says GLM-5.3 uses the same base model as GLM-5.2.
The improvement doesn't come from pretraining scale.
Z.ai says it put more environments, more varied tasks, and more compute into post-training.
So the core of this release is post-training scaling.
How to Read the Coding Gains
Z.ai claims a 50% improvement over GLM-5.2 on its own Code Bench.
On public evaluations, it says GLM-5.3 is the best open model on Terminal Bench 3.0 and Agents' Last Exam.
Internal evaluations and public benchmarks run under different conditions, though.
When you compare numbers side by side, record the model version and run conditions too.
CyberGym Is the Bigger Signal
According to the official post, GLM-5.3 scores 84.5% on CyberGym.
That beats GLM-5.2's 77.2%.
This benchmark measures the ability to find and verify vulnerabilities.
It is not material that teaches anyone how to carry out an actual attack.
In practice, read the result as a sign of what defensive automation could do, and as a warning sign.
Why the Weights Were Delayed
Z.ai says its cyber capabilities grew faster than expected.
So it plans to release the weights after finishing safety evaluations and hardening.
The release is about two weeks after the announcement.
What's confirmed today is a planned release, not weights you can download now.
Before the Weights Ship, Write Down the Conditions
First, track the model card and the weights release as separate items.
Second, note whether each cyber benchmark number comes from an internal or a public evaluation.
Third, if you're considering local deployment, wait for the deployment conditions after the safety hardening.
The GLM-5.3 story isn't only about a performance race.
As open models get stronger, balancing release speed against safety review becomes a key factor in choosing a product.