MAIN STORY
The freedom to run AI, and the loss of control
An open-weight AI model makes its trained parameters available for download. Organisations can run it on their own infrastructure, adapt it to specialised workflows and retain greater control over private data. Chinese developers including DeepSeek, Z.ai and Alibaba have used this approach to grow communities of users and challenge better-funded international rivals.
The trade-off is that safeguards embedded in a downloadable model may be modified or bypassed, while distribution cannot easily be reversed. Closed models, commonly accessed through an application programming interface, can monitor usage and suspend accounts. But their protective filters can also obstruct legitimate defensive research or the analysis of sensitive system logs.
What the security assessments establish
In September 2026, the US National Institute of Standards and Technology’s Center for AI Standards and Innovation described GLM-5.3 as the most cyber-capable open-weight model it had assessed at that point. Its aggregate benchmarking nevertheless placed the model roughly four months behind the leading American frontier in cyber capabilities.

















