We are post training qwen 3.6 and combining it with a custom inference engine and harness to get the most out of a smaller model.