Unsloth Studio with its MTP support: https://unsloth.ai/docs/models/qwen3.6#mtp-guide

Try llama.cpp and Qwen3.6-35B-A3B

Good balance of intelligence and speed.

I think their Max models are far bigger than fits on consumer hardware. People are typically using Apple, AMD Halo, or dGPUs if/when they do smaller versions. Those are all varying degrees of "affordable."