Member-only story
Will the US give in and really ban open weights models? Lets dig in.
Kimi K3 landed in the middle of an argument that is much bigger than one Chinese model: who gets to run frontier AI, where it runs, and whether fear of misuse becomes an excuse to keep that capability behind two American API gates.
The model is not a laptop toy. Moonshot says K3 has 2.8 trillion parameters, a one-million-token context window, native image input, and a sparse MoE design that activates 16 of 896 experts. Its own deployment guidance points to a supernode configuration with 64 or more accelerators. At four bits per parameter, the raw weights alone have a theoretical floor of about 1.4 TB before KV cache, runtime overhead, replicas, networking, or any useful serving headroom.
That hardware reality should end one lazy talking point immediately. K3 is not a model that a random person quietly runs on a gaming PC. It is a serious infrastructure object. But that does not make the opposite answer sensible either: force every company, security team, university, and developer to send their data through a tiny closed-model cartel and hope its policy decisions remain aligned with their work.
I am Caspar Bannink, writing here at CasparAI. I also build HomeScout, an AI rental-search product for…