qwen 3.6 27b and qen35b a3b work like magic, if we get dpark speculative decoding versions of these models it will further improve the throughput