The number that could mislead a buyer evaluating Kimi K3.1 Pro is not its largest one. Moonshot’s model card describes an open-weight mixture-of-experts model with 2.8 trillion total parameters but 104 billion activated parameters. The smaller figure describes active computation, not a complete deployment footprint. For a product team considering hosting, those are different purchasing questions.
A mixture-of-experts architecture routes tokens through selected expert networks rather than using every expert for each token. This card lists 896 experts, 16 selected per token, and two shared experts. Selective activation can reduce computation relative to using the whole expert pool, but the remaining weights still belong to the model a serving system must accommodate. Their placement and movement depend on the implementation. The activated-parameter count alone therefore cannot establish the memory, hardware or cost required to serve it.
Moonshot also lists MXFP4 weights and MXFP8 activations, with quantization-aware training. These lower-precision representations are relevant to deployment, but the supplied excerpt contains no serving configuration or measured hardware requirement. It would be premature to turn the architecture table into a machine count—or conclude that this release makes private hosting affordable for a small specialist business.
Consider a hypothetical document-review product for engineering consultancies. Its builder might want hosting outside a model vendor’s API to control deployment changes. If accommodating the full model requires more infrastructure than one consultancy can justify, a specialist hosting provider could become the practical route. That preserves a possible choice of operator, but shifts continuity questions toward that operator’s capacity, contract and ability to maintain the service. It does not establish privacy or portability by itself.
The useful opportunity is a broader choice of deployment arrangements, not an assumption that open weights make this model economical to operate independently. Hosted access may still be the better decision. Moonshot’s architecture makes the distinction especially visible: selective computation and responsibility for the whole model are separate constraints. For this release, deployment control should be evaluated against an actual serving design and license terms, neither of which is established by the supplied excerpt.
