DSpark-style speculative decoding draft head for Qwen3.6-27B AEON: training recipe, paired eval harness, vLLM patches. +11% serve tok/s vs DFlash (all domains, CI'd). - View it on GitHub