Z.ai released GLM-5.3 on August 14, 2026, positioning it as the strongest open-weights coding model through gains achieved entirely via post-training on the same 743-billion-parameter base as GLM-5.2. The model is available now through Z.ai's API and GLM Coding Plan, with open weights promised in approximately two weeks after security evaluation.
The numbers
GLM-5.3 scored 28.3 on Terminal-Bench 3.0, up from GLM-5.2's 4.6, leading all open-source models on that benchmark. On Z.ai's internal Code Bench, the company reports a 50 percent improvement over GLM-5.2 while using fewer output tokens per task. The model also leads on CyberGym for defensive cybersecurity, though it trails Anthropic's restricted Mythos 5 on offensive exploit benchmarks.
Independent KingBench 3 results showed GLM-5.3 scoring 91.25 percent, matching or exceeding Opus 5, Kimi K3, and approaching Fable 5. The company explicitly designed the model for code auditing and vulnerability discovery, a focus that appears to have strengthened general coding capability as a side effect.
Why it matters
GLM-5.3 is the counter-narrative to Amodei's vision. Where Anthropic pitches trust through proprietary breakthroughs, Z.ai is releasing the weights. In two weeks, operators will be able to download, modify, quantize, and deploy GLM-5.3 without asking permission or paying per token. For founders building AI-native products, that difference is not philosophical—it is architectural. Open weights mean you control the upgrade cycle, the privacy boundary, and the cost structure. That is sovereignty.