RCS-RP-004 — From the Cloud to the Living Room: Consumer Hardware and the Rise of Home AI Agents

Abstract

RCS-RP-004 is the fourth independent research study published by Raiju Cloud System. This working paper examines how consumer computers can support personal AI agents, distinguishing local agent orchestration from local language-model inference. It explains dedicated GPU memory, system RAM, SSD storage, unified memory, quantization, context capacity and the key-value cache. Mac mini and Mac Studio systems are discussed alongside selected AMD and NVIDIA alternatives using explicitly identified external sources. The study combines a focused technical review with a resource-reserved Windows case study using Qwen3-0.6B Q8_0 and llama.cpp on an Intel Core i7-13620H laptop with an NVIDIA RTX 4060 Laptop GPU. CPU-only, partial GPU and full GPU placement were examined at context capacities of 512 and 2,048 tokens. Twenty-four of 27 planned fixed-length requests completed, with 386 resource-monitoring samples. A battery-reserve guard stopped the final configuration. At context capacity 2,048 with a 588-token prompt, median request times were 23.42 seconds for CPU-only placement, 8.34 seconds for partial GPU placement and 0.31 seconds for full GPU placement, with three completed requests per condition. These findings concern one small model on one constrained, battery-powered laptop. They are not unrestricted hardware benchmarks, Mac-versus-PC comparisons or evidence of general autonomous-agent competence. An isolated OmniRoute setup attempt is documented, but routed inference was not benchmarked. The package contains the English research report, a separate machine-readable dataset, reproduction scripts and a SHA-256 manifest. Local measurements, manufacturer specifications, external research and illustrative calculations are explicitly distinguished. This is an independent working paper and has not undergone peer review.