Zhipu AI Releases GLM-5.3, Claims It's the Strongest Open-Weights Coding Model

Zhipu AI has released GLM-5.3, a new version of its coding-focused language model that the Chinese startup claims is the strongest open-weights coding model currently available. The model shares its base architecture with predecessor GLM-5.2, with all reported improvements coming from extended post-training rather than a new pretraining run.

According to Zhipu, the biggest performance jumps appear in agent-based coding tasks. The company has not published specific benchmark scores, parameter counts, context window size, or pricing for GLM-5.3, so these figures cannot be verified at this time.

Cybersecurity focus

One consistent gap between Chinese frontier models and their US counterparts has been cybersecurity capability. Zhipu says it specifically trained GLM-5.3 with data and simulated environments designed for vulnerability discovery. According to Z.ai, the model "began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains" — a claim that has not been independently verified.

Working with security teams in China, Zhipu says the model helped identify 2,436 vulnerabilities across 269 software projects, some of which the company says are up to 40 years old. Zhipu states these findings are documented in a public registry, though the company has not detailed the methodology used to validate each vulnerability or confirm they were previously unknown.

Availability

GLM-5.3 is available now through Zhipu's GLM Coding Plan and is compatible with coding agents including ZCode, Claude Code, and OpenCode. Pricing for the coding plan has not been disclosed in the announcement.

Model weights are set to be released as open source in approximately two weeks, once Zhipu completes internal security reviews. The company has not specified a license type for the eventual open release.

What this means

Zhipu's claim of "strongest open-weights coding model" comes without accompanying benchmark data — no SWE-bench, HumanEval, or agentic coding scores were shared alongside the release. Until Zhipu publishes numbers or third parties run independent evaluations, the claim should be treated as marketing rather than an established result.

The cybersecurity angle is the more concrete development here. If the vulnerability-finding numbers hold up under scrutiny, it would mark a real capability shift for a Chinese open-weights model in a domain where US labs like Anthropic and OpenAI have generally held an edge. The two-week delay before weights ship, explicitly tied to security review, suggests Zhipu itself is treating the exploit-generation capability as something that needs vetting before broad release — a notable admission for a model being marketed on its offensive security skills.

For developers, the practical takeaway is narrower: GLM-5.3 is accessible today only through Zhipu's paid coding plan and select agent integrations, with open weights still pending. Anyone evaluating it against Claude, GPT, or Qwen coding models will need to wait for both the open release and independent benchmark runs to make an informed comparison.