Qwen3.6-27B replaces Claude in local multi-agent test

User ran Qwen3.6-27B via Ollama on a single RTX 3090 for two weeks, replacing Claude as the reasoning layer in a multi-agent orchestrator. The model worked in some areas but failed in others, with trade-offs in reliability and cost.