GLM-5.3: Big Coding Gains From Post-Training Alone

Z.ai released GLM-5.3 on 14 August, and the headline number is its score on Terminal-Bench 3.0, a test of how well a model drives a command line through a long job, climbing from 4.6 to 28.3. What makes that odd is that the brain underneath did not change. GLM-5.3 runs on the exact same base as July’s GLM-5.2, a 753-billion-parameter mixture-of-experts model with a million-token context window. Every gain came from post-training.

This runs against the usual story about model progress. We are trained to expect big leaps from bigger models and fresh pre-training runs that cost a fortune. Z.ai took a model it already had and taught it to be far better at the tasks agents actually get stuck on, refactors, CI triage, and loops that run for hours, without spinning up a new base. On its own internal coding-agent benchmark it scored 50% higher than GLM-5.2.

Table of Contents

How you retrain a model without retraining it

The trick was the training environment. Z.ai used specialised AI agents to generate sandboxes that imitate a developer’s workstation, modelled on real software projects, then had those agents write programming exercises tailored to each one. A separate judge agent checked that every challenge was actually solvable before handing it to GLM-5.3. Some exercises took days to finish, which is precisely what taught the model to hold a plan together over a long horizon rather than sprinting at a single completion. The stack sits on two open-source pieces, slime for moving the model from training to inference, and SAO, an asynchronous reinforcement-learning method that speeds the runs up.

The cybersecurity part is a double edge

Z.ai says GLM-5.3’s security research skill grew faster than it expected as training scaled. On CyberGym, a benchmark for finding code vulnerabilities, it reached 84.5% and beat Anthropic’s Claude Mythos 5, though it fell behind that model on two other security tests. More concretely, the company says GLM-5.3 has already found more than 2,400 vulnerabilities across 269 software projects, about half rated medium severity or worse, including one flaw in code written 40 years ago.

Read that twice. A freely available open-weights model that is genuinely good at locating exploitable bugs is a gift to defenders and to attackers in equal measure. Z.ai is open about the capability outrunning its own expectations, which is honest and slightly alarming in the same breath.

What you can use today

Access is live through Z.ai’s API and its GLM Coding Plan, and it has been rolled out to existing coding-plan subscribers.
The weights are due on Hugging Face under an open licence within two weeks of launch, so around the end of August, when outside testing begins.
The design is unchanged from GLM-5.2, so the million-token context and the mixture-of-experts architecture carry straight over.

Where to keep your scepticism

Most of these numbers are Z.ai’s own. Terminal-Bench and CyberGym are public, but the 50% coding-agent gain is measured on an internal benchmark, and the vulnerability tally has not been independently audited. Until the weights land and outsiders can poke at it, treat the claims as a strong prospectus rather than a verified result. There is also the practical caveat that trails every capable Chinese open-weights model: plenty of Western enterprises will not put customer code near it whatever the benchmark says, and for offensive-security work you would want to know exactly whose sandbox it phones home to.

Still, the lesson is worth banking. If a post-training pass can move a coding benchmark that far on a frozen base, the next year of model progress may be decided less by who has the biggest cluster and more by who builds the cleverest training environment. Z.ai’s answer was to make agents build the gym.

Did you know: the reward signal that shapes a model during reinforcement learning is increasingly generated by other models, so AI now openly grades its own homework.

Sources

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →