M5 Max 36GB powers local AI

36GB unified memory is enough to make local AI genuinely useful, especially if the goal is private, offline assistance for scripts, commands, and routine admin work. It is not enough to make a laptop feel like ChatGPT Pro, but it is enough to replace a lot of “quick question” usage if you pick the right model size.

• –Apple silicon and local runtimes like Ollama and `llama.cpp` are a good fit here; both are built to run efficiently on Apple hardware and support quantized models.
• –Ollama’s own guidance puts 24-48 GiB systems in the 32k-context bucket, which is comfortable for normal chat but still tight for long-agent or heavy coding workflows.
• –Qwen is a reasonable starting point because it ships in multiple sizes; the 7B and 14B classes are the practical sweet spot, while 32B is the upper end of what feels comfortable on 36GB.
• –For IT admin and scripting prompts, local models are usually “good enough” for boilerplate, explanations, and shell snippets, but they still need verification on anything risky.
• –The real win is cost and privacy, not raw quality: if you only need occasional help, local AI can beat a paid subscription on value even when it underperforms the best cloud models.