Even the router got swapped out

Previously, Llama-swap, Open-WebUI, and Claude code router were combined to operate several Qwen series models by role in a 2x RTX 3090 + 1x P40, 128GB RAM environment.

• Qwen 3.5 4B: for semantic routing
• Qwen 3.5 30B A3B Q8XL: general chat, basic document work, web search, large context
• Qwen 3.5 27B Q8XL: more precise responses and simple math/summarization
• Qwen 3 Next Coder 80B A3B Q6_K: code generation
• Qwen 3.5 122B UD Q4KXL / Q6: common sense, debugging, complex reasoning

The problem was the semantic router, as Qwen 3.5 4B often picked the wrong model or ignored override keywords like quick, think, and ultrathink. Even adding lengthy guidance text didn't help, so keywords ended up having to be hardcoded into the router script.

Another weakness was excessive thinking token consumption by the 27B model. It tended to use a lot of tokens even on simple problems like PEMDAS, and the post cuts off right as it moves into the next comparison.