Qwen3.8-27B at 256K on a 24GB RTX PRO 4000 SFF (432 GB/s): 50 tok/s with MTP

A researcher successfully ran Qwen3.8's 27B model with a full 256K token context on a single 24GB GPU, achieving 50.44 tokens per second through careful optimization of quantization, speculative decoding (MTP), and CUDA kernels rather than relying on individually "best" components. The experiment demonstrated that thoughtful system engineering and the fit between multiple factors—quantization method, drafter precision, memory layout, and workload—matter as much as raw model size or hardware specifications for local LLM inference performance.