2026-08-29 / Signal #3
A Chinese open-weight model just claimed the coding-and-hacking crown
“The new open model isn’t just good at writing your code. It’s getting suspiciously good at breaking other people’s.”
The Story
Z.ai released GLM-5.3 as open weights; it uses the same base as GLM-5.2 but post-training jumps made it SOTA among open models on coding benches (Terminal Bench 3.0, DeepSWE) plus unexpected cyber/vulnerability-discovery gains. 1M context, reasoning always on, already on Ollama and local stacks.
Why It Matters
The “open-source is catching up” story just got a cyber-capable, long-horizon coding agent you can actually download; builders now have a free-ish frontier coder whose extra skill is finding exploits.
Evidence
Hugging Face model card (zai-org/GLM-5.3), Hacker News (719 points), Z.ai docs.
Sources
Daily scan: 2026-08-29