I ran a 27B AI Model on a Mac mini M6… Here's What Happened! | Bonus Storage Hack

Can a Mac mini M6 really run today's best 27B parameter model locally? And what happens when a USB4 external SSD turns out to be FASTER than Apple's internal storage? I unboxed the Mac mini M6 (32GB RAM, 256GB SSD), ran Qwen 3.8 27B locally using Apple's MLX framework, pushed its context window to 64K tokens, and tested a clever storage upgrade that could save you $$$. No cloud. No API fees. No subscriptions. Just local AI, real benchmarks, and a few surprises along the way. 🔍 WHAT I DISCOVERED 🧠 27B AI on a Mac mini ~11.5 tokens/sec running Qwen 3.8 27B at 3.5-bit DWQ quantization Three quantized versions tested: speed vs quality Why memory bandwidth, not raw compute, becomes the bottleneck Why I decided against speculative decoding 🎯 The 64K Context Challenge Can a locally running AI model find a hidden sentence buried in 61,000 tokens? I put Qwen's long-context capabilities to the test, with real results. 💾 BONUS: The Storage Hack Apple's upgrade: A$300 (~US$211) for just 256GB more storage My alternative: A$349 (~US$245) for a 1TB Corsair USB4 SSD External SSD read speed: 4,069 MB/s Mac mini internal SSD read speed: 2,631 MB/s Yes, the external SSD delivered faster sequential reads in my test! I also tested running my local AI setup from external storage. ⏱️ CHAPTERS 0:00 Intro 0:23 Mac mini M6 unboxing 4:34 Running out of space, and a clever fix 7:11 Running Qwen 3.8 27B locally 9:27 Why memory bandwidth is the limit 11:02 Testing three quantized versions 11:51 Checking quality, not just speed 12:46 The 64K context test 14:58 Moving everything to the SSD: real results 16:44 The verdict 17:18 Your turn: what works on your Mac? 17:39 Next: one photo to a game-ready 3D model 🧰 MY SETUP Apple Mac mini M6 — 32GB unified memory, 256GB SSD (A$2,049 / ~US$1,441) Corsair EX400U Survivor 1TB USB4 SSD (A$349 / ~US$245) AI Model: Qwen 3.8 27B, 3.5-bit DWQ quantization Inference Framework: Apple MLX Everything runs locally, without cloud inference. All benchmark numbers are from my own tests. Results may vary depending on model configuration, storage setup and workload. US$ prices are approximate, using 1 AUD = 0.703 USD (25 Oct 2026). 🚀 COMING NEXT Can this same Mac mini turn a single photograph into a game-ready 3D model, completely offline? In my next video, I explore turning a photo into a 3D ski boot for my game, without relying on cloud AI. Subscribe to follow the experiment! 💬 LET'S COMPARE RESULTS Are you running local LLMs on a Mac mini, MacBook or another Apple Silicon machine? What model are you using, how many tokens/sec are you getting, and have you found a way to make speculative decoding worthwhile? Share your setup and results in the comments. I'd love to compare notes! #MacMiniM6 #LocalLLM #AppleSilicon #MLX #OfflineAI #Qwen #USB4 #LocalAI