Z.ai's GLM-5.3 Claims the Open-Weights Coding Crown, But Its Weights Aren't Public Yet

Z.ai released GLM-5.3 on August 14, 2026, claiming it is the strongest open-weights coding model the company has measured — achieving frontier-level agentic coding benchmark results with roughly 750 billion parameters, about a third the size of Moonshot AI's 2.8-trillion-parameter Kimi K3.

What Actually Changed From GLM-5.2

Z.ai says GLM-5.3 keeps the same base model as its predecessor, deriving every capability gain purely from scaled-up post-training rather than a larger or retrained base. The company reports its steepest gains in coding and, notably, in cybersecurity capability — which it says grew faster than anticipated as training scaled, a detail worth flagging for any security team tracking dual-use model capability.

A Genuinely Crowded Field, Not a Two-Horse Race

• DeepSeek moved V4-Pro to general availability on August 13, one day before GLM-5.3's launch
• Kimi K3's full weights shipped July 27, already establishing itself as the world's largest claimed open-source model
• Z.ai's own comparison table benchmarks GLM-5.3 directly against DeepSeek-V4 Pro, Kimi K3, and OpenAI's proprietary GPT-5.6 Sol across coding, cyber, and agentic task suites

One important caveat for anyone evaluating GLM-5.3 today: the weights are not yet public. Z.ai says it needs roughly two more weeks for safety evaluation and hardening before release — so current claims rest on the company's own benchmark reporting, not independent verification.