Making Qwen Faster with ThinkingCap

Reasoning models are dominating the current field of conversational AI, and among them, Qwen 3.6 27B is one of the most popular open-source models. But if you look at what the model reasons about when you simply say "Hi," you'll see hundreds of tokens spent analysing the user's input, identifying their intent, and selecting the correct response strategy — and all of that just for a simple reply like "Hello! How can I assist you today?" In this talk, we'll walk you through our journey of modifying this model to cut the number of tokens spent during the reasoning phase by half on average, without affecting performance. It's an approach that's clearly struck a chord - the model and its quantisation variants have crossed 500,000 downloads on Hugging Face.