The cover of this piece was made on my own laptop. My dictation goes nowhere either: I hold a shortcut, speak, and the text appears without anything being uploaded. Meanwhile a scheduled task runs overnight on Anthropic’s servers while my laptop is shut. That is not inconsistency but a per-task trade-off, and the rule of thumb that falls out of it is short: local when the input is sensitive or latency matters, cloud when it has to keep running while you’re away.
The three approaches side by side
| Approach | Where your input stays | What it costs | What you get | What broke for me |
|---|---|---|---|---|
| Fully local | Never leaves your device, not even your prompt | Disk and battery: around 1 GB for a speech model and 3.5 GB for the cleanup, plus hardware that can handle it | No meter, no service that can decide tomorrow your style is no longer allowed, and it works without internet | A half-installed environment, a component that worked depending on which folder I was in, and a model that went silent instead of returning an error |
| Fully cloud | On the provider’s server, together with your instruction | A monthly subscription, with the heaviest features starting on the priciest plan | It keeps running while you’re offline, and you install nothing | The usage ceiling you hit on a productive day, and features that only trickle down to cheaper plans later |
| Cloud control, local execution | Your files stay on your machine, the control path goes through a service | Usually included in the same subscription, no separate price | You can start something from your phone that works on your own material | With your machine off, nothing happens: it moves the controls, not the execution |
What to weigh
How sensitive is the input? This is the first question, not the last. For a one-off cover it hardly matters. For work that isn’t public yet, or for dictation where you mention a colleague’s name halfway through, it matters completely. Speaking leaks more than typing: the hesitation, the aside, the sentence you start over.
Does the delay matter? Voice that goes above seven hundred milliseconds starts to feel sluggish. Cut the network out of the chain and a chunk of that delay falls away without you optimising anything. For a task that runs overnight, latency is irrelevant.
Is anyone waiting on it while you’re elsewhere? Then local is out. A model on your laptop does nothing with the lid closed.
What does it cost, and who does that tie you to? A subscription is predictable, but it is also a dependency: the price, the limits and the feature split change without you having a say. Running locally asks for hardware and evenings, but the model on your disk keeps doing what it did.
What running locally really costs
Generating is not the problem. The image model I use, Bonsai Image, has been compressed far enough to fit on a laptop and makes an image in about six seconds on an M4 Pro. It even runs on a phone.
The time sits in the toolkit around it. I wrote earlier about why I do it this way anyway: because I own it, because it lets me hold one consistent style, and because you only understand a tool once you have been stuck with it. That last part is the point, not a side effect.
Dictation sits differently. FluidVoice runs both the speech model and the cleanup locally, and it was that second step I long thought impossible. Raw dictation is messy, and most apps send that mess to a server to turn it into tidy sentences. That is exactly where the most sensitive text sits.
What the cloud buys
One thing, and it is a big one: it keeps running while you’re away. When Cowork moved off the desk that also changed what it is useful for, and that piece sets out why a task which executes itself overnight is a different class of tool from one that waits on you.
The price is that the work moves to someone else’s machine. For a news briefing that is fine. For material I don’t own, it is the wrong place.
The caveat: open source moves the dependency, it doesn’t remove it
Anyone running locally easily assumes they are now independent. Not quite. The dictation app is GPLv3, but the models doing the work come from NVIDIA, from Cohere, from OpenAI. Free to use, but not built by me and not maintained by me. The dependency has shifted from a service to a set of model weights, and that is a gain, not a liberation.
There is an access side too. Local is not free: it costs disk, battery and hardware that is good enough. Anyone on an older laptop simply cannot make this choice, which makes “just run it yourself” advice that only works for those who can afford it.
And one specific to local apps: to type into any text field, such an app asks for access to your keyboard and your microphone. An app you pulled off GitHub then gets to watch permanently. Running locally does not make those permissions smaller, only easier to justify.
There is a reason I’m sharp on this. By day I work on digital sovereignty, on the question of who owns which data. It would be odd to argue that and then act differently in the evening.
Frequently asked questions
Is running locally always safer?
For your input, yes: it never leaves your device. But a local app often needs access to your keyboard and microphone, which is a real risk you don’t have with a web service. Safer on one axis does not mean safer on all of them.
Do I need expensive hardware to run a model locally?
For images, not necessarily any more: some models have been compressed enough to run on a laptop and even a phone. For fast, good results recent Apple Silicon does help. Budget several gigabytes of disk per model.
If you could only run one thing locally, what would it be?
Dictation. That is where the input is most sensitive, because speaking makes you say more than you would type, and where the delay is most noticeable. Image generation is the most fun to run yourself, but rarely the most urgent.
So is the cloud worse?
No. For work that has to keep running while you’re offline there is no local alternative. The question is not which of the two is better, but which fits this task.
Figures and model names come from my own pieces on this setup, written between June and August 2026. Cloud pricing and limits change regularly; check them with the provider before relying on them.