Source: Decrypt — 2026-08-14
Summary
Z.ai released GLM-5.3 on the same roughly 744-billion-parameter base model as GLM-5.2, with every capability gain coming purely from additional post-training rather than any architecture or scale change. Terminal-Bench 3.0 score rose from 4.6 to 28.3, DeepSWE v1.1 climbed from 46.2 to 66.9, and on CyberGym — a benchmark for finding and validating real vulnerabilities from source code — GLM-5.3 scored 84.5%, edging out Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). Open weights and broader API access are being staged over roughly two weeks after launch, pending a safety review, after Z.ai says the model's vulnerability-finding capability progressed faster during training than the company expected.
Key Takeaways
- GLM-5.3 uses the identical base model as GLM-5.2 — Z.ai attributes the entire capability jump to post-training (reinforcement learning across a large, diverse set of task environments), not architecture or parameter-count changes.
- Terminal-Bench 3.0 score jumped from 4.6 to 28.3 (more than 6x), DeepSWE v1.1 rose from 46.2 to 66.9, and Agents' Last Exam improved from 23.8 to 28.5.
- On CyberGym, a benchmark measuring whether a model can find and validate real vulnerabilities given source code, GLM-5.3 scored 84.5% — narrowly ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%).
- ExploitBench score more than doubled, from 24.4% to 54.4%, and Z.ai says the model flagged 2,436 candidate vulnerabilities across 269 open-source projects during evaluation, with 1,097 rated high-severity or critical.
- Z.ai says it added vulnerability-discovery training environments expecting incremental gains but instead saw capability progress further along the exploitation chain than planned, and is staging open weights and broader API access over roughly two weeks post-launch pending a safety review.
Reel Script
Hook: Z.ai just took the exact same 744-billion-parameter model it shipped months ago and, without touching the architecture, made its coding benchmark score jump more than sixfold — proof that in 2026, post-training alone can be the difference between a mediocre model and a state-of-the-art one.
Core Concept: Most headline AI model jumps come from scaling up — more parameters, more pretraining data, a new architecture. GLM-5.3 is the opposite case: Z.ai took the identical base model underneath GLM-5.2 and pushed nearly all of the improvement through post-training, the reinforcement-learning stage where a model practices completing tasks in simulated environments and gets rewarded for succeeding, rather than the earlier stage where it just learns to predict text. That distinction matters because it's a much cheaper lever to pull than retraining a base model from scratch, and the size of the jump — a benchmark score moving more than sixfold — shows how much headroom current base models apparently still have if you just train them harder on the right tasks. One of those task categories was vulnerability discovery: giving the model real source code and rewarding it for finding and proving a genuine security flaw exists. Z.ai says that specific training ingredient produced a side effect it didn't fully plan for — the model's offensive security capability, the ability to find exploitable bugs, improved faster than the team expected, which is exactly the kind of dual-use result that makes a pre-release safety review matter.
Hands-On: The benchmark table is the single most screen-worthy artifact here, because every number moved in the same direction and none of them came from a bigger model. Terminal-Bench 3.0, which scores a model on completing real terminal-based coding tasks end to end, went from 4.6 to 28.3 — more than a sixfold jump. DeepSWE v1.1, a software-engineering benchmark, rose from 46.2 to 66.9. Agents' Last Exam climbed from 23.8 to 28.5. And ExploitBench, which scores exploit-generation capability specifically, more than doubled, from 24.4% to 54.4%.
The CyberGym result is the one worth a closer look, because it's a three-way photo finish at the top: GLM-5.3 scored 84.5%, Claude Mythos 5 scored 83.8%, and GPT-5.6 Sol scored 83.6% — a spread of under one point separating the top three models on a benchmark that specifically measures whether a model, handed real source code, can find and validate a genuine vulnerability. Behind that headline number, Z.ai says GLM-5.3 flagged 2,436 candidate vulnerabilities across 269 real open-source projects during its evaluation runs, with 1,097 of those rated high-severity or critical — not synthetic test cases, but flaws in software people actually run.
That vulnerability-discovery result is also why you can't download GLM-5.3's weights yet. Z.ai is releasing model access through its GLM Coding Plan and ZCode first, and staging open weights plus broader API access over roughly the two weeks following launch, gated on a safety review of exactly the capability the benchmarks just demonstrated.
Takeaway: The real story isn't that GLM-5.3 topped a leaderboard — it's that a model lab can now buy sixfold benchmark gains without touching pretraining, which resets expectations for how fast open-weight models can catch up to frontier labs. Watch the safety review closely: how Z.ai handles staged weight release here will likely become the template other labs copy.