LM Studio steers M4 Max users to Qwen3-Coder

36GB unified memory is the sweet spot where local coding stops being a toy problem and becomes a real workflow. Qwen3-Coder-30B-A3B wins here because it keeps active parameters low while leaving enough headroom for context and repo-scale work.

• –`Qwen3-Coder-30B-A3B` is a 30.5B MoE model with only 3.3B active weights, so its runtime footprint is much friendlier than the raw parameter count suggests.
• –LM Studio supports the model in both GGUF and MLX, and the Mac-friendly MLX path is the right way to squeeze more speed out of Apple Silicon.
• –For pure coding, a specialist model beats a generalist one at this memory tier; `Qwen3.5-35B-A3B` is the fallback only if you want more breadth than code focus.
• –The real constraint on a 36GB Mac is not whether a model fits, but how much context and KV-cache headroom you can keep without slowing the machine down.