Qwen 3.5 27B optimization boosts M2 Max speed

The "dense model tax" is hitting Mac users hard as they trade MoE speed for the superior coherence of large dense models. Dense models activate every parameter per token, whereas MoE counterparts like the 35B-A3B version only activate ~3B, making them inherently 5-10x faster. Setting "Max Concurrent" to 4 in LM Studio is a common speed killer; dropping this to 1 prioritizes single-stream bandwidth for the 27B parameters. Weight quantization is critical; running a Q8 weight quant on a 27B model pushes the M2 Max bandwidth (400 GB/s) to its limit, whereas Q4_K_M offers a 2-3x speedup with minimal logic loss. macOS limits wired memory by default; running sudo sysctl iogpu.wired_limit_mb can unlock the full 64GB for the GPU, preventing the thrashing that causes "heartbeat" job failures. Transitioning to the MLX framework often yields 20-40% better performance on Apple Silicon compared to standard GGUF backends.