Z.ai debuts GLM‑5.

Chinese AI developer Z.ai released its open‑source LLM GLM‑5.3, setting new benchmarks in coding and cybersecurity research.

Chinese AI firm Z.ai unveiled its latest open‑source large language model, GLM‑5.3, touting breakthroughs in long‑horizon code generation and fortified capabilities for cybersecurity research.

Key advancements in coding

GLM‑5.3 extends its context window to 64 k tokens, allowing developers to feed entire codebases and receive coherent, multi‑file suggestions without losing track of earlier dependencies. The model also integrates a specialized “debug‑first” training regime that prioritises error detection and automated patch generation.

Early benchmarks show GLM‑5.3 outperforming its predecessor, GLM‑5, on the HumanEval suite, achieving a pass rate increase of several percentage points while maintaining comparable inference latency.

Cybersecurity research upgrades

In the security domain, Z.ai equipped GLM‑5.3 with a curated corpus of vulnerability disclosures, exploit code, and threat‑intel reports. This enables the model to generate realistic attack simulations, assist in penetration‑testing scripts, and propose mitigation strategies in natural language.

The company emphasized that all training data were sourced from publicly available repositories and that the model includes built‑in safeguards to refuse instructions that facilitate malicious activity.

Open‑source release and community support

GLM‑5.3 is released under the Apache 2.0 license on GitHub, accompanied by a comprehensive toolkit for fine‑tuning on domain‑specific datasets. Z.ai also announced a bounty program to encourage contributions that improve the model’s safety and performance.