想要流畅、游刃有余，不是加点配置就行，而是要换一个级别的机器：从 Mac Studio M5 Ultra 起步。
Smooth, effortless local AI isn't a matter of adding a few options — it takes a different class of machine. Start at Mac Studio M5 Ultra.
决定"能不能开"。模型得整个装进油箱才能上路，装不下就根本启动不了，或者一边开一边去远处加油，慢得没法用。
Decides whether you can drive at all. The whole model must fit in the tank; if it doesn't, the car either won't start or keeps detouring to refuel — too slow to use.
决定"开得多快"。AI 每说一个字，都要从油箱里抽一次油。管子越粗，字蹦得越快。
Decides top speed. Every word the AI writes draws fuel from the tank. A wider line means words come out faster.
决定"起步多快"。你丢给 AI 一大段代码或文档，它得先全部读完才开口。发动机小，起步就要干等很久。
Decides acceleration. Hand the AI a big codebase or document and it must read it all before replying. A small engine means a long wait before it moves.
流畅的体验要三样都够：油箱装得下、油管供得上、发动机拉得动。Mac mini 三样都卡在"刚刚够"的边上。
A smooth experience needs all three: a tank that holds it, a line that keeps up, an engine that pulls. The Mac mini sits right at "barely enough" on all three.
Mac mini 像小孩，Mac Studio M5 Ultra 像健壮的成人。给他们背同样重的书包（同样大的模型），小孩也背得起来，但走得慢、走不远、一路喘气；成人背着同样的包，步子轻快，还能腾出手多拿东西（更长的对话、更多的代码、同时开几个工具）。
The Mac mini is a child; the M5 Ultra Mac Studio is a strong adult. Give both the same heavy backpack (the same model). The child can lift it, but walks slowly, not far, out of breath. The adult carries it easily and still has hands free for more — longer conversations, more code, several tools at once.
这就是"能跑"和"游刃有余"的区别。我之前在 16GB Mac mini 上的经历就是这样：技术上能跑，但慢到抓狂，实际上用不了。
That's the difference between "it runs" and "it runs effortlessly." My experience on a 16GB Mac mini was exactly this: technically it ran, but so slowly it was unusable in practice.
油箱刚好装满，没有余地。Mac mini 最多 64GB。一个 80B 级模型就要占掉约 46GB，剩下的还得分给系统和对话记忆；120B 级模型（约 63GB）则根本装不下。
The tank is full with nothing to spare. The Mac mini tops out at 64GB. An 80B-class model alone takes about 46GB, and the rest must cover the system and conversation memory; a 120B-class model (about 63GB) doesn't fit at all.
输油管只有 Ultra 的四分之一。Mac mini 是 307GB/s，M5 Ultra 是 1.2TB/s，差了将近 4 倍。
The fuel line is a quarter of the Ultra's. 307GB/s on the Mac mini versus 1.2TB/s on the M5 Ultra — nearly 4× apart.
发动机小，起步慢。丢一整个项目的代码给它，Mac mini 要干等很久才开始回答。写代码、读长文档这类真实用法，最常卡的就是这一步。
A small engine means slow starts. Hand it a whole project and the Mac mini waits a long time before answering. Real work like coding or long documents stalls most often right here.
| 机型Model | 油箱（最大内存）Tank (max memory) | 油管（带宽）Fuel line (bandwidth) | 大陆起售价CN starting price |
|---|---|---|---|
| Mac mini M5 Pro | 64GB | 307GB/s | ¥12,999 |
| Mac Studio M5 Max | 128GB | 460–614GB/s128GB 需选高配 GPU128GB needs the top GPU | ¥19,999 |
| Mac Studio M5 Ultra | 96 / 256 / 512GB | 1.2TB/s | ¥46,99996GB / 1TB |
| 模型（体积）Model (size) | Mac mini 64GB | M5 Max 128GB | M5 Ultra 96GB | M5 Ultra 256GB |
|---|---|---|---|---|
| Qwen3-Coder-Next 80B~46GB | 能跑，但慢Runs, slowly | 顺畅Smooth | 游刃有余Effortless | 游刃有余Effortless |
| gpt-oss-120b~63GB | 装不下Doesn't fit | 能跑，速度一般Runs, average speed | 快，但油箱偏满Fast, tank nearly full | 游刃有余Effortless |
M5 Max 128GB 油箱够大，但油管和发动机都只有 Ultra 的一半左右，属于"装得下、开不快"。
The M5 Max 128GB has a big enough tank, but its fuel line and engine are roughly half the Ultra's — it fits, but doesn't go fast.
买本地电脑像买车，用云端 AI 像打车。四五万到近十万元，够打很多次车；而且本地能跑的开源模型，能力明显比不上最顶尖的云端模型——相当于买来的车比出租车档次低。车还会折旧，明年又出新款。
A local machine is like buying a car; cloud AI is like taking taxis. ¥47k to nearly ¥100k pays for a lot of rides. And the open models you can run locally are noticeably weaker than the best cloud models — the car you buy is a class below the taxi. Cars depreciate, and next year brings a new model.
所以只有三种情况买车更划算：数据不能出门（隐私）、经常在没网的地方用（离线）、或者每天打车打得特别多（用量极大）。
So buying only wins in three cases: data can't leave the building (privacy), you often work without internet (offline), or you ride so much every day that fares pile up (very heavy usage).
起步：Mac Studio M5 Ultra / 96GB，¥46,999。80B 级模型游刃有余；120B 级能跑得很快，只是油箱偏满，对话太长会吃紧。
Entry: Mac Studio M5 Ultra / 96GB, ¥46,999. Effortless for 80B-class models; fast on 120B-class too, though the tank is nearly full and very long conversations get tight.
推荐：M5 Ultra / 256GB，一次到位。120B 级模型也游刃有余，还能同时开多个工具。美国售价 $9,499，国行价格以官网选配为准。
Recommended: M5 Ultra / 256GB, done once. Effortless even on 120B-class models, with room for several tools at once. US price $9,499; check Apple's CN configurator for the local price.
不推荐：Mac mini 和 M5 Max。Mac mini 是小孩背重包，能走但走不动；M5 Max 油箱够大，但油管只有 Ultra 一半，跑不出流畅感。
Not recommended: Mac mini or M5 Max. The mini is a child with a heavy pack — it moves but can't keep going. The M5 Max has the tank but only half the Ultra's fuel line, so it never feels smooth.
买之前再确认两件事：一是先想清楚要跑哪个模型，模型越大越要往 256GB 走；二是 M5 Ultra 的性能数字目前来自苹果官方，等首批用户实测出来再下单更稳。
Two checks before buying: first, decide which model you'll run — the bigger it is, the more 256GB makes sense; second, M5 Ultra performance numbers currently come from Apple, so waiting for early real-world tests is the safer move.
一句话：想要流畅、游刃有余，Mac Studio M5 Ultra 96GB 起步，预算够就直接 256GB。
In one line: for smooth, effortless local AI, start at Mac Studio M5 Ultra 96GB — go straight to 256GB if the budget allows.