M5 Max MacBook Pro sparks model debate

This is less a “best model” question than a reality check on local agentic coding: the M5 Max has enough headroom to make serious on-device inference practical, but tool-calling quality and latency still decide the winner.

• –Qwen 3.6 27B gets the strongest vote because it balances coding quality, long context, and quantization resistance better than the bigger but clunkier alternatives
• –Gemma 4 31B is competitive, but commenters still flag tool-call reliability as a recurring weakness for agent workflows
• –MiniMax M2.7 is the dark-horse pick for people who care about autonomous coding behavior over raw benchmark vanity
• –Apple’s M5 Max pitch fits the discussion: the machine is no longer the bottleneck for many local coding stacks, the model and runtime are
• –The practical takeaway is to optimize for agent loop quality, not just parameter count; a smaller model that follows instructions cleanly will beat a larger one that drifts