A developer has implemented RadianceMemory and RadianceSwap, enabling KV-cache paging to host RAM with byte-identical output via sparse attention and in-process multi-model serving that dynamically moves VRAM between models at runtime. The work was validated on Qwen3.8-Flash-Next running across two AMD Radeon AI PRO R9700 GPUs (gfx1201, tensor parallelism=2) using forks of the Radiance inference engine. A 23-page paper with measurements is available on GitHub.