June access restrictions around frontier models made a familiar dependency visible. Whether framed as export controls, access policy or risk management, the effect is the same for a team relying on a remote model: a capability can become unavailable outside its own technical roadmap.
Kimi K3 and Inkling make the counterpoint concrete. Kimi describes K3 as an open 2.8-trillion-parameter model and has said its full weights will be released by 27 July. Thinking Machines released Inkling with full weights available for people to customize. These are not interchangeable systems, nor do they erase the compute, data, talent or infrastructure needed to use them. They do widen the set of actors that can inspect, adapt and deploy a capable model.
On 17 July, China called for open source, openness, collaboration and sharing in AI. That message sits alongside an environment in which access to leading systems is increasingly shaped by national-security decisions. The tension is not rhetorical: it concerns who may build on a model, where it may run and whether critical capability can be withdrawn.
Open weights are therefore part of digital sovereignty, not a complete substitute for it. They can reduce a single provider's control over access and support local adaptation, auditability and continuity. The strategic question is not simply whether a model is open or closed, but which dependencies a jurisdiction, company or institution is prepared to accept when frontier AI becomes leverage.
Sources