Status: 草稿（积压 #26）| 2026-09-02 | 目标平台: Dev.to / Medium | 联动: 方案 A（本地小模型工作流模板）、gig #2（模型选型评估） Three data points in three days tell a story that's easy to miss if you're watching only the frontier labs.
1. Apple got caught off guard. Last week's Hacker News thread on AI demand for Mac Mini and Mac Studio (287 points, 334 comments) — Apple reportedly under-supplied because people are buying Macs specifically to run models locally. Not to browse. Not to code. To run inference.
2. A 104GB model on a 48GB Mac. Yesterday's Show HN: running Qwen3.8-Flash-Next at ~12 tokens/sec on a 48GB Mac Mini (138 points). The gap between "model too big for this hardware" and "model runs fine, slightly slow" is closing with quantization and better runtimes. 12 tok/s isn't ChatGPT-fast, but it's private and free per token.
3. Local setups are becoming routine. A second post the same day: "My local model setup on an M4 Pro Mac Mini" — no longer a novelty, just a setup note. When something stops being impressive enough to argue about, it's becoming infrastructure.
The frontier labs compete on the biggest models. That's a war you don't need to fight. The local tier is different: it's about fit, not scale — which model runs on which hardware, at what speed, with what quality tradeoff. That's a knowledge problem, not a compute problem. And knowledge problems are where small operators win.
Three concrete gaps worth building for:
When hardware sells out because of AI workloads, the software layer around that hardware is still empty. That's the gap. The frontier is crowded; the local tier is not — and it's getting real faster than the headlines suggest.
~500 words. Sources: HN 287pts (Apple Mac Mini/Mac Studio AI demand, 8/31), 138pts (Qwen3.8-Flash-Next on 48GB Mac, 9/2), 16pts (M4 Pro local setup, 9/2), 565pts (small transformer beats LLMs, 9/2).