Local model or subscription?

A question that comes up on X most weeks, in some form: "Should I buy a Mac with 64 GB of memory for AI, or just keep paying for my chat subscription?" The replies split into two camps. Local models are free and private and subscriptions are a rip-off, or the reverse. Neither reply answers it, because the real question is three questions: what do you gain, what do you give up, and under which assumptions does the choice make sense? I come back to the 64 GB Mac at the end.

What you are actually comparing

"Local versus cloud" hides two comparisons. The first is hardware: a model on a laptop you already own costs almost nothing extra, while a machine bought for the purpose has a price tag. The second is product: a subscription like Claude Pro comes with search, file uploads, a phone app, memory, and integrations. A local setup is a model plus an application, and how much the application covers varies.

Also: open weights is not open source. You can download the weights, but the license may still restrict what you do with them, and the training data is almost never included. Read the license first.

Start with what you need

What will you do with it: quick questions, coding all day, or 200-page contracts? How often, and for how many people? What matters most: accuracy, speed, confidentiality, or convenience? And is tinkering part of the fun? If so, the hours are still spent, but the learning may be worth them to you. Put that in the calculation.

Count retries and correction time

Test on your own tasks, in your own language, on your own subject. Look for specific failures: does it follow instructions, invent facts, write code that runs, lose the middle of a long document? Then ask whether a smaller model is enough.

Measure the whole interaction. A model that answers in three seconds and needs two corrections loses to one that takes twelve seconds and gets it right. Count retries and time the corrections, because a confidently wrong answer gets no retries and still costs you.

The model is not the product

When someone says a local model is "as good as Claude", the comparison is usually about the text of a chat answer. Then they upload a PDF, ask about yesterday's news, or dictate a message while walking. The product gap is easier to see: list the features you use without thinking, then check which ones your local application provides. LM Studio, for example, ships document chat and tool integrations. Coverage varies by application; the rest you add and maintain yourself.

Hardware, briefly

Memory is often the first constraint. A model quantized to 4 bits, meaning weights stored at lower precision to save space at a possible cost in quality, needs roughly 0.5 to 0.6 GB per billion parameters. So an 8B model needs about 5 GB, a 32B about 20 GB, a 70B about 40 GB, for the weights alone. Add working memory for the conversation, which grows with context length, plus headroom for the operating system and your browser. A discrete GPU needs the model in its own memory; Apple Silicon shares one pool. LM Studio recommends at least 16 GB of RAM and says 8 GB Macs can manage smaller models with modest context. Sustained inference also drains a laptop battery fast; measure it on your own machine.

Count the whole price

The usual assumption is €1,200 of hardware versus €20 a month, which gives 60 months. That is where the calculation starts.

break-even months = net upfront cost ÷ (monthly subscription spending avoided − monthly local running cost)

Net upfront cost is the extra you spend over what you would have bought anyway, minus the resale value attributable to that extra. A €1,200 machine bought purely for this is €1,200. A €300 memory upgrade to a laptop you needed anyway is €300.

Subscription spending avoided is the term people get wrong. Local use only saves subscription spending if it reduces what you actually pay: cancelling, downgrading, avoiding extra usage charges, or subscribing for fewer months. Same plan at the same price means zero.

Running cost is mostly electricity, and only the extra draw counts. A hypothetical laptop pulling an extra 50 W for 3 hours a day, 30 days a month, at €0.30 per kWh, uses 4.5 kWh and costs €1.35. A desktop GPU pulling an extra 250 W on the same schedule costs €6.75.

So: a €1,200 dedicated machine that replaces a €20 plan and costs €5 a month to run gives 1,200 ÷ (20 − 5) = 80 months. A €300 upgrade replacing the same plan at €1.35 a month gives 300 ÷ 18.65 = about 16 months. If the subscription stays as it was, the denominator is zero or negative and there is no financial break-even. Local is then justified by other reasons or not at all. Count your time too, at a rate you believe rather than your billing rate.

Date every price. Claude Pro is $20 per month billed monthly, or $200 billed annually, shown as about $17 per month, excluding tax, on claude.com/pricing as of 6 September 2026. Check current pricing before deciding. The euro figures above are illustrative, not conversions.

Speed, privacy, control, and working offline

Which setup is faster depends on hardware, model, and task. Measure time to the first word and time to a finished, correct answer. Only the second pays the bills.

Local can keep everything on the device. Can, not does. A search plugin, a sync feature, or a cloud fallback sends content out again. Ollama documents this well: it runs locally, listens only on the local machine by default, and has a switch to disable its cloud features. If you expose the model server to your network, you also need to secure access.

Two local advantages get less attention than they deserve. Local works on a plane and during an outage. And a local model behind a local API can grind through repeated jobs, with no per-token charge and no rate limit. A consumer subscription is an interactive product; the hosted equivalent for automation is an API, priced per token.

Local also lets you freeze a working setup: the model you validated in March is the same in October. A subscription changes under you, usually for the better, sometimes not.

Three people, as starting points

These are illustrative configurations, not verdicts. Pick the nearest one, run a week of your own tasks on it, and let the result decide.

A casual user with a 16 GB laptop. A reasonable first try is a 4-bit model in the 4B to 8B range. If it handles their rewriting and questions with few corrections, they lose nothing by using it. If it fails on long documents and recent events, a free hosted tier already covers that, and local is worth it only when confidentiality or curiosity outweighs the missing features.

An enthusiast with a 32 GB Mac or a 24 GB GPU bought for other reasons. Try 14B to 32B models. The incremental cost is near zero, so the only question is whether the results hold up. If correction time on hard tasks stays below what $20 a month is worth to them, local can be the default. If not, a hybrid, with the extra cost and the extra decision per task counted.

A professional working six hours a day under client confidentiality. Being a professional does not set a minimum model size; the tasks do. Test the largest model the existing hardware runs against the hosted option on a week of real work. If the local model costs more than an hour or two of extra correction a month, the hosted plan or API wins on money. Local stays for the material that cannot leave the building, whatever the cost.

Back to the 64 GB Mac

Before buying, write down a week of your actual tasks and run a 32B-class model on them, on a rented machine or at a hosted open-weights provider, noting retries and correction minutes against your current subscription. Then work out the extra cost of 64 GB over the Mac you would buy anyway, and decide what you would really do with the subscription: cancel, downgrade, or keep. If the model holds up and the break-even is shorter than the time you expect to keep the machine, buy it. If not, buy the Mac you needed anyway and keep paying. That is a less quotable answer than either camp gives, which is probably why it never gets many likes.