M1 Max 64GB finds LLM sweet spot

This is the right question: on Apple Silicon, the best experience usually comes from a well-quantized 26B-35B model with a solid runtime, not from forcing a frontier-size model into memory.

• –64GB unified memory is enough to run serious local assistants, especially with Q5/Q4 quantization and longer contexts, so the machine is not the blocker
• –Gemma 4 26B A3B is the likely comfort pick for chatty, low-friction use; Qwen3.6 35B A3B should be stronger on reasoning and broader tasks but will feel heavier
• –llama.cpp and MLX/oMLX are the relevant Mac runtimes here, and the main tradeoff is speed versus context length rather than raw “can it load” capacity
• –For coaching and project management, instruction-following and conversation quality matter more than coding benchmarks, so the user should optimize for tone and consistency