Kimi K3.1 Pro Released as Open Weights, Raising Questions on Self-Hosting Costs

Moonshot AI has published the weights for Kimi K3.1 Pro, a Mixture-of-Experts model with 2.8 trillion total parameters, 104 billion active per token, and a context window above one million tokens. Because only a subset of experts runs for each token, inference can be more efficient than a comparably sized dense model, but the full checkpoint of roughly 1.5 TB still must be stored and served. Documented vLLM deployment recipes rely on high-end accelerator setups such as NVIDIA B300 or GB300 systems and multi-node Zhenwu configurations, and a consumer GPU cannot host the complete model. The article cautions that third-party hardware cost estimates are not official quotations and that developers should verify supported configurations, KV cache capacity, throughput, and networking before investing. It concludes that deciding between self-hosting and using an API depends on production requirements beyond parameter counts.