Handwriting read on your own machine.

What are the three local backends?

Whichever one is in use, a page is drawn to a temporary image in your system's own temp directory, read from there, and deleted straight afterwards. Apple's recognition is invoked through a program that ships with macOS and opens no connection of its own.

Which model should I load on my own server?

The plugin suggests none, for the reason in the next section. But the question people ask is which one to load, so here is a measurement rather than an opinion: the same fifteen pages the nightly reads, one Mac, every model in its 4-bit build through the plugin's own engine.

Load the 8B wherever it fits: it is far the best of them. The 2B is the one to try on 8 GB. Skip the 4B — it reads no better than the 2B and needs nearly twice the memory, so there is no machine it is the answer for. The names are what LM Studio and Ollama call them.

The 8B and 2B rows are from 16 September 2026 and the per-page figures behind them are in the repository, beside the prompt, the image renderer and the engine build they were measured with; a check there refuses a build of the plugin whose pins have moved past a figure, so a number here is either current or the build does not pass. The 4B was measured once, on 10 September, and is not kept current — the plugin never offers it.

One machine, one hand. Take it as a starting point for what to load first, not as a promise about your notebook.

Why does the plugin not pick a model for me?

Apple's recognition aside, the plugin suggests no model to you — not on a local server, not on a cloud service. A model id shipped inside a plugin goes stale the day its provider retires it, and the failure you get then reads like a broken key rather than what it is. So you take a current name from your provider's own list. Until you do, a sync says plainly that no model is set, instead of quietly producing nothing.

Two things the settings screen will tell you about a model, where the server can be asked at all. A text-only model cannot look at a page image, so it will never work. And a model that thinks out loud before answering is far slower and can drop a page altogether — its answer runs past the length it is allowed, and that page is left out rather than saved half-read. The sync reports that at the end, and the fix is a model that does not reason.

What does local recognition cost me?

• On Windows with an Intel or AMD chip, and on Linux, nothing is read until you set a server up. Notes still arrive, with the page drawn as you wrote it and no text under it. That is the honest gap, and it is why the local server route exists: it is the one that works everywhere.
• The downloadable model is a real download, and which one you get depends on the machine. A Mac with 16 GB of memory or more is offered the 8-billion model: 6.2 GB to fetch, about 8 GB of memory while a page is read, and roughly eight seconds a page against Apple's fraction of one. A Mac with 8 GB is offered a 2-billion model instead — 1.6 GB to fetch, 2.9 GB while it works, and 6.6 % character error on the same fifteen pages, which is less than half of what Apple's own recognition makes of them. On Windows on Arm the same two start at 24 GB and 16 GB. Nowhere else is it offered.
• Where those floors come from. Each model was measured at its peak while it read a page, four more gigabytes are left for everything else on the machine, and the total is rounded up to a size machines are actually sold in. The 8B peaked at 7.99 GB, which gives 16 GB; the 2B at 2.89 GB, which leaves 114 MB of headroom on an 8 GB Mac and is the thinnest margin of any of them. The Windows on Arm numbers are deliberately not derived from the Mac ones: the only Windows figure anyone has measured is the older model's, on a processor-only path, and scaling it by a ratio would be inventing a measurement rather than making one.
• Taking a newer model costs the disk twice, for as long as you want it to. The one you already downloaded is not deleted when the new one arrives — it is the way back if the new one disappoints — so both sit there until you say otherwise. A button in settings removes the old one when you no longer want the space.
• Windows on Intel and AMD is excluded on purpose. The engine behind that model has been flagged by Windows Defender as malware for years — wrongly, but nothing this plugin does to someone else's unsigned binary changes what a scanner decides about it. Windows on Arm uses a different build, which is not flagged.
• Sending a page to a cloud service is possible, and it is the one exception. Four such services are part of Pro, the optional add-on. They run on an API key of yours, they are off until you switch them on, and that is the only case in which an image of your page leaves the machine.

Does my handwriting leave my machine? →

How does it misread, and how do I spot it?

No handwriting recognition is exact, and these fail in ways worth telling apart. Apple's mistakes usually look wrong on the page: a word comes out as letters that are not a word, and your eye catches it. A model's mistakes read as fluent, ordinary prose that simply is not what you wrote, which is much harder to notice.

Both are now measured on the same fifteen pages, which they were not before. On one Mac on 10 September 2026 the model a new install downloads read them at 1.8 % character error, and Apple's recognition at 17.4 %. So the download buys roughly ten times fewer wrong characters, and costs a few seconds a page against Apple's fraction of one.

If you installed the plugin before that model existed, you have its predecessor, which reads the same pages at 4.3 %. It keeps working and nothing replaces it behind your back — the settings screen offers you the newer one with both figures beside each other, and downloads it only if you ask.

Two things that pair is not. It is one machine and one set of pages, in my handwriting — yours will differ, and Apple's recognition alone moved by nearly two points between two Macs running the same version of its engine. And a smaller build of the same model family is a different model: the table above says what the 2- and 4-billion versions read like, and neither figure is a discount on the 8B's.

Which is why the page you wrote is embedded directly above its text in the note. Checking a line means glancing up, not opening the tablet.

The nightly measurement, page by page →
What the tablet’s own conversion does instead →

How accurate is the transcription? →

When is this not for you?

• You are on Windows with an Intel or AMD chip, or on Linux, and you do not want to run a server. Then take the page renders and leave the text alone — they are still searchable by title, tag and folder.
• You need a guaranteed accuracy figure for someone else's compliance form. There isn't one: it depends on your handwriting and on the backend you chose.
• You want the text to be right without checking it. Nothing on the market does that yet, and a page that claimed otherwise would be selling you something.

It is the default, not an upgrade.

All three local backends are in the plugin itself and cost nothing. Pro adds cloud services for anyone who wants them — €24 once, and off unless you turn them on.