Don't Buy a Mac Mini for Local AI Until You Pick the Model

The new Mac mini runs from $899 to $2,699, and every step buys room for bigger AI models and speed. So which step do you need? That depends on the model, so I ran the models on the same M5 Pro chip and timed every answer. In this video: what the B in a model name means, why memory bandwidth sets the speed, why the faster Qwen model is the worse coder (five real coding runs each), where the minutes go, and what each Mac mini configuration actually runs, from 16 to 64 GB, with prices and speeds. Chapters 0:00 The Mac mini page 0:28 Two models, one guess 1:16 What the B counts 2:07 What more knobs cost 3:20 Why memory speed sets the pace 3:53 M6 or M5 Pro: the number that counts 4:31 The answer 4:55 Five runs each 5:49 Dense and experts 7:09 Where the minutes go 8:24 What your Mac decides for good 8:38 Picking your Mac mini, 16 to 64 GB 10:42 How I would choose 11:16 Next time How I measured: MacBook Pro, Apple M5 Pro (15-core CPU, 16-core GPU, the same chip as the $1,699 Mac mini), 48 GB, 307 GB/s. LM Studio, 25 September 2026. Qwen 3.6 in GGUF Q4_K_M, Gemma 4 in MLX 4-bit. Coding job: five runs per model with Qwen's recommended settings (temperature 0.6, top_p 0.95, top_k 20), 16,000-token cap, 12 unit tests. Memory: peak wired memory while answering a 3,400-token and a 28,000-token prompt. M6 speeds are estimates: my measured speed × M6 bandwidth ÷ 307, rounded down. The rule landed within 10% on my machine (Gemma 4 31B: limit 16.7, measured 15.2 tokens a second). Prices: US, base storage, September 2026. The GPU share macOS allows by default is two thirds of memory up to 32 GB and three quarters above; it can be raised with the iogpu.wired_limit_mb setting. Sources Mac mini tech specs (bandwidth, cores): https://www.apple.com/mac-mini/specs/ Prices and memory upgrades: https://daringfireball.net/2026/08/co... Qwen3.6-27B model card (SWE-bench Verified 77.2): https://huggingface.co/Qwen/Qwen3.6-27B Qwen3.6-35B-A3B model card (73.4; 256 experts, 8 routed plus 1 shared): https://huggingface.co/Qwen/Qwen3.6-3... Gemma 4 model card (LiveCodeBench v6): https://ai.google.dev/gemma/docs/core... Part 2, why local AI takes minutes to answer: Why Local AI Takes Minutes to Answer (I Ti... Part 3, what AI a 16 GB Mac can actually run: What AI Can a 16 GB Mac Actually Run? (I S... Narration uses an AI clone of my own voice.