Who should not switch to local AI

I sell a local AI app, so treat everything here accordingly. But I'd rather you read this and decide correctly than download something, find it doesn't do what you assumed, and conclude that local AI is a toy. Most of the disappointment I see comes from a mismatch that was predictable in advance. Here's how to predict it.

You need things that are structurally impossible offline

Some capabilities aren't a maturity problem. They cannot work on a machine with no network by definition, and no amount of engineering changes that.

• Live information. Today's news, current prices, whether a flight is delayed, what a company announced this morning. A local model knows what was in its training data and nothing after. If most of your questions are about the present, you need something with web access.
• Anything that touches an external service. Reading your inbox, checking a calendar, querying a live database, calling an API on your behalf. Offline means offline.

People sometimes expect that a good enough local model would eventually do these. It won't. That's the wrong axis.

You need capabilities we specifically don't have

Separate category: things that are possible locally, just not in our app today. I'd rather name them than let you find out after downloading.

• Image generation. We don't do it. That's a diffusion model, a different architecture entirely. We can read an image you give us; we can't make one. If you want both on the same Mac you'd run a separate tool alongside.
• Voice mode. No speech in, no speech out.
• A phone app. Mac only. If you want AI in your pocket, this isn't it.
• Cloud sync across machines. Your conversations live on the machine that made them, which is the point, and also a limitation.

Your hardware doesn't fit

This one is arithmetic, and it's the most common cause of a bad first experience.

We're Apple Silicon only — M1 or later, macOS 12 or newer. Intel Macs are out. Windows and Linux are out. And within Apple Silicon, RAM decides everything:

If you have 8 GB and were hoping to run a 27B model, the answer is no, and no setting fixes it. Buying a new Mac to save $20 a month is also, obviously, terrible arithmetic. Stay on cloud.

You're working at the top of the difficulty range

Here's where I have to be careful not to oversell, because our own measurements set the boundary.

On everyday tasks, our local 27B is genuinely strong — in an execution-verified evaluation it was effectively 100% correct on common coding problems, and 83% on a deliberately hard novel set. Against Claude Opus on a 54-prompt head-to-head it matched on 98.9% of rubric checks. Those are good numbers and I stand behind them.

But on a blind slice of SWE-bench Verified — real bugs in real repositories — the same model measured about 45%. That's a real number, not a leaderboard claim, and it's below what the strongest cloud coding agents achieve. If your work is the hard end — novel research reasoning, long autonomous multi-file agent runs, problems where the last few percent of capability decides the outcome — the frontier cloud models are still ahead, and I'd rather you use them than be quietly disappointed in us.

You don't want to think about it

Cloud AI has one genuine, underrated advantage: there's nothing to choose. You open it, it's the model, it works.

Local AI asks you to make decisions. Which tier fits your RAM. Whether this question needs the fast small model or the slow big one. Whether a 14 GB download is worth the disk. We work hard to make those defaults sensible, but the decisions exist. Some people find that interesting. Some find it friction they never asked for. If you're the second kind, that's a completely reasonable preference and cloud is the better product for you.

Two bad reasons to switch

Because local AI doesn't hallucinate. It does, at roughly the same rate, for exactly the same reasons. Where it runs has nothing to do with whether it's right. We caught one in the act and wrote it up.

Because it's obviously cheaper. Sometimes. Our Pro is $20/mo or $149/yr against ChatGPT Plus at $20/mo — real but not dramatic, and the free tier is genuinely free. If you'd have to buy hardware to make it work, the maths inverts completely.

So who is it actually for?

Briefly, since this page is about the other direction: people who handle material they'd rather not upload, people who hit usage caps and resent them, people who work somewhere with bad connectivity, people who want the tool to keep working regardless of a vendor's pricing decisions — and people who just like owning their tools. If none of those is you, you're better served elsewhere, and I'd rather say that than take the download.