Chinese AI startup Zhipu AI released GLM-5.3 on August 14. The company calls it the strongest open-weights coding model available today. The claim rests on Zhipu’s own benchmarks.

GLM-5.3 shares the same base architecture as its predecessor, GLM-5.2. All of the reported gains come from extended post-training alone. Zhipu says coding capability jumped 50 percent over the prior version.

The model carries 743 billion parameters. Zhipu trained it with data and environments built specifically to find software vulnerabilities. The company says GLM-5.3 uncovered 2,436 vulnerabilities across 269 projects during testing. Some of those flaws dated back as far as 40 years.

GLM-5.3 is available now through the GLM Coding Plan. It works with coding agents including ZCode, Claude Code and OpenCode. The model weights themselves are not yet downloadable. Zhipu says it plans to release them in about two weeks, once security reviews are complete.

The release adds to a crowded week for AI model launches. Google shipped Gemini 3.7 Flash on August 13, and other labs have pushed out competing coding models in recent days. Zhipu, also known as Z.ai, has positioned itself as a low-cost, open-weights alternative to Western labs. That strategy has helped Chinese AI firms gain traction with developers who want to inspect or modify model weights directly.

Zhipu is one of several Chinese AI startups collectively nicknamed the country’s AI tigers, a group that also includes Moonshot AI and MiniMax. These firms have leaned into open-weights releases to build developer mindshare beyond China, since most leading American labs keep their strongest models closed. GLM-5.3’s coding focus puts it in direct competition with DeepSeek’s V4-Pro and Alibaba’s Qwen models. Both rivals have also pushed coding-specific updates in recent months.

Security researchers are likely to test Zhipu’s vulnerability-detection claims once the weights become public. Until then, the numbers come solely from the company itself. That gap between claim and independent verification is common across the fast-moving open-weights model market.