I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.