Alibaba disclosed new safeguards for its Qwen4-Max AI model, including excluding training material intended to improve offensive hacking. It said the model scored 14.7% on a secure-code-generation benchmark, up from 8% for its predecessor, and reduced average task costs by around 60% in internal testing.
Alibaba said Qwen4-Max retains cybersecurity knowledge useful for detecting and fixing software vulnerabilities while excluding material intended to improve offensive hacking capabilities. The company also said the model had stronger safeguards against chemical, biological, radiological and nuclear threats, and that internal evaluations found no evidence it could significantly enhance users’ ability to carry out dangerous tasks.
Alibaba said Qwen4-Max scored 14.7% on a benchmark measuring secure code generation across functions, files and software repositories, compared with 8% for its predecessor. Built with 2.4 trillion parameters, it reduced average task costs by around 60% in internal testing, the company said. Alibaba plans to release smaller open-weight models from the Qwen4 family for developers to download and modify.