Hardware: RTX 5090 | Model: Qwen3.6-27B | Framework: BeeLlama.cpp Full benchmark scripts, raw data, config, and generated artifacts are available on request — just DM or comment below. I spent the last week benchmarking DFlash speculative decoding combined with KV cache compression strategies on Qwen3.6-27B. The results are surprising enough that I wanted to share them for anyone running local inference. Setup GPU: NVIDIA RTX 5090 (32GB VRAM) Model: Qwen3.6-27B in two qu...