The Pulse
Z.ai and Concordia Set Out Six Stages for Open-Weight AI Risk
Z.ai and Concordia AI published a framework for managing open-weight AI risks, organized around six stages from risk identification to governance. It adds societal resilience as a fourth dimension for assessing risk and describes measures i
AI.info Team · Written with AI ·
Z.ai and Concordia AI published the Frontier Open-Weight AI Risk Management Framework, a guide for developers who release model weights and the wider ecosystem that downloads, modifies and runs them. Concordia AI says the framework was first announced on September 21, 2026, at the UN Digital Cooperation Day.
The framework starts from three challenges of open-weight release: a release cannot be undone, safeguards can be removed through fine-tuning, and no single party can monitor use or withdraw the model. The authors describe the document as a companion to the Frontier AI Risk Management Framework 2.0 and adapt its risk-management approach for releases where developers no longer control deployment.
“The Framework does not resolve every tension between openness and safety.”
The authors of the Frontier Open-Weight AI Risk Management Framework
A companion to Framework 2.0
The new framework follows six stages: risk identification, risk thresholds, risk analysis, risk evaluation, risk mitigation and risk governance. At each stage, it considers which protocols from the earlier framework rely on developers controlling access or monitoring use, and proposes adaptations for open-weight releases.
The framework is intended for open-weight general-purpose models with capabilities in risk-relevant areas approaching or matching the frontier, rather than the broader open ecosystem. Its contributors are Brian Tse, Oskar Galeev, Weibing Wang, Yuan Cheng and Liang Fang of Concordia AI, and Wenbin Qiao, Xiaochen Wang and Chen Zhang of Z.AI.
A fourth dimension for thresholds
The earlier framework analyzed risk through three dimensions: deployment environment, threat source and enabling capability. The open-weight framework adds a fourth: societal resilience, or society’s capacity to absorb and respond to AI-enabled harms. Concordia’s description says this extends the analytical lens of the earlier framework.
The four dimensions cover where a model operates, who might misuse it, what capabilities it enables and how well society can withstand resulting harm. The report’s approach recognizes that open-weight models can be modified after release, so assessing risk involves more than examining a model in its original form. The framework sets out risk thresholds and calls for analysis at different points in a model’s development and release process.
Risk evaluation and mitigation
In its evaluation stage, the framework sorts models into green, yellow and red risk zones. These correspond, respectively, to full release, restricted or staged release, and suspension. For yellow-zone models, the report describes options such as staged release or structured access rather than treating the choice as simply open or closed.
The report cites Z.ai’s GLM-5.3 as an example of staged access: vetted security partners evaluated the model in controlled settings before broader access and API availability, with full weights released after safety evaluations. It also points to measures that can be applied before release or remain relevant afterward, including curation of training data early in development.
The framework’s six-stage process concludes with mitigation and governance. Together, the stages address risks from identifying potential harms through to setting oversight and accountability for models whose weights can circulate beyond their developers’ control.