GLM-5.3 Goes Open-Weight
Z.AI dropped GLM-5.3 on Hugging Face yesterday — open weights, Apache-adjacent license, ready to run locally. This is not a paper announcement or a blog post about a model they'll release later. The weights are live.
The headline: GLM-5.3 uses the same base as GLM-5.2. Every gain comes from post-training. And the gains are not marginal.
- Terminal Bench 3.0: 28.3% (vs GLM-5.2's 4.6%). That's not an improvement — that's a regime change.
- DeepSWE: 66.9% — competitive with closed frontier models like GPT-5.6 Sol (72.7%) and Fable 5 (69.7%).
- CyberGym: 84.5% — state of the art for vulnerability discovery, beating GPT-5.6 Sol (83.6%) and Fable 5 (83.8%).
- ExploitGym (6h): 130 passes — up from 39 on GLM-5.2. More than 3x on exploitation benchmarks.
The cyber capability angle is the weird one. Z.AI says it emerged faster than they expected as they scaled post-training. The gains are largest further up the exploitation chain — not just finding vulnerabilities, but exploiting them. This is a trajectory worth watching.
On the infrastructure side: GLM-5.3 runs on SGLang, vLLM, Transformers, Unsloth, and notably Ascend NPU (Chinese hardware). The reasoning_effort parameter lets you dial thinking budget between low/high/max. 128K output tokens, 1M context window in some benchmarks.
The open-source coding model frontier just shifted. GLM-5.3 is not a curiosity — it's slotting in above every other open-weight model on coding and agentic benchmarks, and touching closed frontier territory. The fact that it runs on Chinese hardware and the cyber capability curve is accelerating makes this more than a model drop. It's a signal.