The article documents a pragmatic shift away from chasing large language models on constrained hardware (RTX 3070, 8GB VRAM). Instead of repeatedly failing to run 35B or 12B models with degraded performance, the author runs two smaller models simultaneously via llama.cpp on different ports: Gemma 4 E2B (2B, for chat/knowledge) and FluentlyQwen3-Coder-4B (4B community fine-tune, for coding tasks). Both fit entirely in GPU memory with context windows intact, accessed via browser tabs. The practical win isn't that these small models are secretly excellent—the author is clear they're modest—but that they remain responsive and usable without constant monitoring, context ceilings, or CPU offloading slowdowns. Total VRAM footprint stays comfortably below 8GB with KV caches. The claim is testable: two 2GB-class models with reasonable context beat one 8GB+ model that forces painful trade-offs on modest consumer hardware. The source doesn't provide benchmarks comparing latency or quality, only workflow observations.