Moonshot AI's Kimi K3, a 2.8-trillion-parameter, open-weight, 1-million-token-context model that mixes KDA-based linear attention with periodic full-attention layers, is set to ship its full open weights by July 27. Ahead of that, the vLLM project published a preview of the engineering work going into serving it efficiently once the weights land, and the details are a useful window into what it actually takes to self-host a frontier-scale open model on day one rather than waiting weeks for the ecosystem to catch up. The core problem vLLM is solving is that Kimi K3's linear-attention architecture doesn't play nicely with the prefix caching techniques that make serving repeated or similar prompts cheap in standard transformer models, because KDA compresses history into a fixed-size recurrent state rather than keeping a full KV cache. vLLM's fix separates the physical size of that KDA state block from the granularity used for prefix matching, so the server can still get useful partial cache hits without needing to store recurrent state at every tiny attention boundary. On top of that, the preview describes fused kernels for KDA decoding, projections, and convolution, a reimplemented MLA module, and MXFP4 mixture-of-experts execution work, plus NVIDIA kernel tuning and an early AMD implementation using a custom MoE kernel. For teams planning to self-host large open-weight models instead of paying API prices, this kind of inference-engine groundwork is often the real bottleneck, since a model can be open the day weights are published and still be impractical to run efficiently for weeks. vLLM shipping this work in lockstep with the model's release lowers that barrier and is a good signal to watch if K3, or its architecture pattern, becomes a template other labs copy.