Kimi K3's Weights Are Free — but Running Them Takes a Data Center

We already covered Kimi K3's open-weight release from China's Moonshot AI — 2.8 trillion parameters, the largest open model to date. But an open license doesn't mean the model is actually accessible to run for just anyone. That openness has an infrastructure price tag, and it's what actually determines who benefits from the open-model race between China and the U.S.

• Even compressed to 4-bit MXFP4 format, the weights take up roughly 1.4 terabytes of memory — before accounting for runtime buffers and cache.
• Moonshot recommends a configuration of 64 or more accelerators on a single high-bandwidth interconnect; independent estimates put the realistic minimum at around 18 Nvidia Blackwell-class GPUs.
• Companies that can't afford to build a cluster like that will still turn to Moonshot's API or cloud providers — recreating the exact vendor dependency that open weights are supposed to eliminate.

Open weights, in this size class, do not push capability out to the edge.

"Open weights, in this size class, do not push capability out to the edge," Techi's analysis notes, describing the release as a deliberate move in the contest over whether the most capable models are things you rent or things you hold.

An independent assessment from infrastructure platform Northflank confirms it: no source suggests the full model can run on a single GPU, and the practical deployment path is reserved cloud capacity, not owned hardware. In other words, Kimi K3 expands China's presence in the open-model race symbolically — as a public fact of openness — but not technically, for the overwhelming majority of developers.

None of this should be read as personalized investment advice.