Qwen 35B-A3B MoE runs ~4x faster than 27B dense locally

In local coding-maintenance tests on an R9700/llama.cpp setup, Qwen 35B-A3B MoE hit ~116 tok/s vs ~30 tok/s for the 27B dense model (~3.9x faster), while the coding-quality gap was much smaller than expected.