Guide
A recent Mac can run a vision model locally. Answers are often slower and less exact than a large cloud model's, and it costs disk space and memory. Ollama is the better pick if you want an MIT-licensed runner with a local API, or if you own an Intel Mac. LM Studio is the better pick for a desktop app with MLX on Apple silicon. Dayflow fits only if you want a record of your workday.
You drag a screenshot into a local model, type “what is on this screen?” and wait. The fans spin up. The first reply takes long enough that you wonder if something broke. When it arrives it’s half right. It spotted the browser and the spreadsheet but misread the column headers, named the wrong app, and made up a line of text that isn’t there.
Or you haven’t got that far yet. You’ve read that you can run AI on your own Mac and keep your screenshots private, and now you’re looking at a model download page that wants more disk space than you have free, with no idea whether your laptop can handle it.
This article covers what a local vision model on a Mac does well, where it struggles, and a handful of low-tech ways to find out whether it fits your Mac and your work before you commit a weekend to it. Disclosure up front: Dayflow, a Mac app that can use a local model to describe your workday, is our product. It comes up once, near the end, along with where it’s the wrong pick. Facts about Ollama and LM Studio come from their own documentation and pricing pages as checked on October 4, 2026.
Most of it comes down to one tradeoff. When the model runs on your Mac, your Mac does all the work. The large cloud models people compare against run on hardware far bigger than a laptop. Running locally gets you privacy and independence from a service. You pay for that in waiting time, memory and sometimes accuracy.
A few things make screenshots especially hard:
None of this means local vision is broken. It means it rewards narrow questions, clear images and a bit of checking.
What to do: On paper, finish this sentence: “I want a model to look at ___ and tell me ___.” Be specific about the input and the output, and list two or three things that would count as a wrong answer.
Illustration: “I want a model to look at screenshots of my work and tell me which project I was on. A wrong answer names the wrong client or calls a code editor a text document.” That’s testable. “Understand my screen” isn’t.
How to tell after a week: Read the sentence again. If you can say whether the model did that job, the step worked. If you keep moving the goal (“well, it’s good at other things”), tighten the sentence and start over.
What to do: Open the Apple menu, choose About This Mac and write down four things: the chip (Apple M series or Intel), the macOS version, the amount of memory, and the free disk space (System Settings, then General, then Storage).
Then compare against what the vendors publish:
Illustration: Someone with an Intel MacBook and very little free disk learns from this five-minute check that LM Studio is out and that Ollama would run on the CPU only. That’s worth knowing before an evening spent downloading.
How to tell after a week: You haven’t hit a “disk full” warning, and you haven’t had to delete something you needed to make room for a model.
Both tools load an open model and let you send it a picture. They differ in how they’re built and what you pay for.
Ollama is MIT licensed and runs models through a CLI, a desktop app and a local API, with Python and JavaScript libraries. That makes it the natural pick if you want to script things. It also offers optional cloud-hosted models that need no local download. LM Studio is a proprietary desktop app that runs models with llama.cpp on every platform and with Apple’s MLX on Apple silicon Macs. It says it can run fully offline once model files are downloaded.
| Ollama | LM Studio | |
|---|---|---|
| Mac support | macOS 14+, M series CPU and GPU, Intel CPU only | Apple silicon only, macOS 14.0+ |
| License | Open source (MIT) | Proprietary |
| Free plan | $0: run models locally, with starter cloud usage credits | $0: local LLMs using llama.cpp and MLX |
| Paid plans | Pro $20 / mo. or $200/yr billed annually; Max $100 / mo.; Team $500 / mo. (early access); Enterprise custom | Bionic+ $20USD / month; Pro $100USD / month |
| What paid plans buy | Cloud usage credits | Cloud inference and web search |
| Memory guidance | None published; see the docs | 16GB+ RAM recommended on Mac |
You don’t need a paid plan with either to run models on your own Mac.
Illustration: Someone who already lives in the Terminal installs Ollama. Someone who wants a chat window and a model browser installs LM Studio. Neither installs both “to compare” on day one.
How to tell after a week: You can load a model, send it a screenshot and get an answer without looking anything up. If you’re still fighting setup by day five, switching runners is a reasonable call.
What to do: Take ten screenshots of your real work: a busy browser, a spreadsheet, code, a video call, a cluttered desktop. Ask the same question about each. On paper, mark each answer right, partly right or wrong, and note what it got wrong.
Illustration: Over ten screenshots, a person might find the model reliably names the app and the general task but misreads document titles and gets numbers in tables wrong. That’s useful. It tells you to ask “which app and what kind of work?” and not to rely on it for figures.
A plain sheet works fine for this. The daily work log template is handy if you want to record what you were actually doing next to what the model said.
How to tell after a week: Run the same ten screenshots again with the same question. If your scores are stable and you can predict where it will fail, you understand the model well enough to use it.
What to do: Use a kitchen timer or your phone’s stopwatch. Time the first answer after opening the runner, then the third answer in the same session. Do it once with your usual apps open and once with heavy apps closed. Write the numbers down.
Illustration: Someone notices the first answer takes noticeably longer than later ones, and that everything slows down while a video export is running. They decide to batch their questions instead of asking one every few minutes.
Keep an eye on battery too if you’re on a laptop. Sustained model use is real work for the chip. The note on screen recording and battery covers how to think about background load on a Mac.
How to tell after a week: You know roughly how long to wait, it no longer surprises you, and you’ve chosen when to run the model around that.
What to do: List on paper what you’d never want to leave your Mac: client documents, health records, private messages. Then confirm the runner is set to match.
Both tools add cloud options on top of local use. Ollama says it doesn’t see your prompts or data when you run locally. Its cloud-hosted models do process prompts and responses, though the vendor says it doesn’t store or log that content. You can disable cloud features entirely with the OLLAMA_NO_CLOUD=1 setting. LM Studio says it can operate entirely offline once models are downloaded. Its paid plans add optional US-hosted cloud inference, which the vendor describes as Zero Data Retention.
Illustration: A consultant who handles client spreadsheets turns off Ollama’s cloud features on day one, so a cloud model can never be picked by accident.
How to tell after a week: You can say, without checking, whether each question you asked stayed on your Mac.
What to do: For anything you’ll act on or share, put the answer and the screenshot side by side. Check names, numbers and any quoted text. Treat the model’s output as a first draft.
Illustration: A model summarizes a screenshot as “reviewing the Q3 budget.” The window actually shows Q2. Thirty seconds of checking catches it before it ends up in a status update.
How to tell after a week: You catch errors before anyone else does, and you’ve stopped being surprised by the kinds of mistakes the model makes.
Many people who want a local vision model want it for one job: knowing what they actually did all day without writing it down. That’s narrower than general image questions, and it’s what Dayflow does.
Dayflow is a free, MIT-licensed Mac app for macOS 14 or later. It stores full screen recordings and a local timeline on your Mac and uses AI to turn them into a readable account of your day. You don’t have to remember to start a timer or rebuild the afternoon from memory. If that’s the question you’re stuck on, where did my day go shows the idea.
Its Local AI mode connects to what this article has covered. A supported model runs on your Mac through Ollama or LM Studio, and screenshot analysis stays on-device. Setup and local inference use your own hardware, so steps 2 and 5 still apply. If your Mac struggles, you can configure your own cloud provider, whose charges and terms apply, or use Dayflow Pro at $20 a month, where Dayflow manages cloud analysis. The local and cloud processing page spells out each mode. You can export a date range as Markdown from Settings → Export. AI summaries can miss or misinterpret work, so step 7 applies here too.
Dayflow is our product, and Dayflow publishes this article.
For more on what it stores and where, see is Dayflow safe.
A local vision model on a Mac is real and useful, but it’s not a free version of the biggest cloud models. Expect to wait, to give up disk space, and to check what it says, especially small text and numbers. Find out cheaply first: write the job down, check About This Mac, pick one runner, and score ten of your own screenshots on paper. If the job turns out to be “show me what I did today,” that’s the one case where a purpose-built record is worth a look. For everything else, Ollama or LM Studio on its own is the simpler path.
Yes, with limits. Ollama runs on macOS Sonoma (14) or newer and uses the CPU and GPU on Apple M series chips, but only the CPU on Intel Macs. LM Studio needs Apple silicon and macOS 14.0 or newer. Expect slower, less precise answers than a large cloud model, and run a few of your own screenshots through it before you rely on it.
It depends on how you work. Ollama is MIT licensed and offers a CLI, a desktop app, a local API, and Python and JavaScript libraries, so it suits scripting. LM Studio is a proprietary desktop app that runs models with llama.cpp and also with Apple's MLX on Apple silicon. Both are free for local use. If you can't decide, try one for a week on your own screenshots.
No. LM Studio's system requirements say the Mac version needs Apple silicon and macOS 14.0 or newer, and Intel Macs are not supported. If you have an Intel Mac, Ollama is the option to look at. It runs on x86 Macs, but only on the CPU, so expect it to be slower than on an Apple silicon machine. Check the vendor docs for current details.
LM Studio recommends 16GB or more of RAM on a Mac and says 8GB machines may work with smaller models. Ollama's current docs give no RAM minimum. An older rule of thumb about memory per model size is no longer on Ollama's pages, so don't treat it as official. For a specific model, read the model's page in the Ollama docs, then test it on your own Mac.
Not when you run models locally. Ollama says it does not see your prompts or data when you run locally. Its optional cloud models do process prompts, and you can turn cloud features off with OLLAMA_NO_CLOUD=1. LM Studio says it can run fully offline once model files are downloaded. Its paid plans add optional US-hosted cloud inference, which the vendor describes as Zero Data Retention.
Running models locally is free in both. Ollama's paid plans are Pro at $20 / mo. or $200/yr billed annually, Max at $100 / mo., Team at $500 / mo. (early access) and custom Enterprise, and they pay for cloud usage. LM Studio lists Bionic+ at $20USD / month and Pro at $100USD / month, which pay for cloud inference and web search, not for running local models.
Ollama's documentation describes models as tens to hundreds of GB in size. On macOS it keeps them at ~/.ollama/models, so that folder can grow fast if you try several. Before you download anything, check free space in System Settings, and decide in advance which test models you will delete once you've picked one.
No. For one-off questions about an image, use Ollama or LM Studio directly. Either lets you send a picture to a model and read the answer right away. Dayflow is built around an ongoing record of your workday, with full screen recordings and a timeline stored on your Mac. That's more setup than you need for a single question.
With Local AI selected, a supported model runs on your Mac through Ollama or LM Studio, and screenshot analysis stays on-device. Recordings and the timeline are stored on your Mac either way. One caveat: if you connect an assistant to a cloud service, that conversation can receive the Dayflow data it retrieves. Local storage doesn't make a cloud assistant conversation local.
Screenshots pack small text, icons and overlapping windows into one image, and many models scale images down before reading them. Fine print, code and dense tables are the first to suffer. Cropping to the area you care about, using a larger window or zoom level, and asking narrow questions usually gets better answers than asking a model to describe everything.
Dayflow is $0 and open source. Full recordings stay on your Mac, with local AI available for on-device analysis.
Download for MacOS macOS 14+ · MIT licensed · no account required