Offline OCR on a Mac: The One You Already Have vs the One You'd Install

Your Mac has done on-device OCR for years. Here is what Apple's Vision framework actually does, and the narrow case where an OCR model earns its gigabytes.

Flat illustration of a page shape beside a column of short horizontal bars on a cream background
9 min read
1,895 words

Two OCR-specific models trended on Hugging Face within a day of each other this week, both small enough to run on a laptop, and the obvious question is whether to install one. The less obvious question, and the better one, is whether you need to: offline OCR on a Mac has shipped with the operating system for years, it runs entirely on the machine, and for most of the paper in front of you it is the right answer.

What follows is what Apple’s Vision framework actually does, the specific rules that trip people up, and the narrow set of documents where a model earns the gigabytes.

The offline OCR you already have on your Mac

Apple’s documentation is unambiguous about where the work happens: “all of Vision’s processing happens on the user’s device to enhance performance and user privacy”. No account, no upload, no network. That single sentence settles the privacy question that sends most people looking for a local model in the first place — and it is a stronger guarantee than several of the paid apps in the local AI file search wave manage to state about themselves.

Vision offers two recognition paths, and they are genuinely different pieces of machinery rather than a quality slider.

Apple documentation diagram comparing the Fast path, which does character detection then character recognition, with the Accurate path, which does neural-network text detection then recognition, both feeding NLP-based language correction
Apple's Vision documentation, captured 31 Oct 2026: the Fast and Accurate paths differ before they converge on the same language-correction step.

Fast or Accurate, and why it matters

The Fast path finds individual characters, then uses a small machine learning model to recognise those characters and words. Apple’s own description is that this “is similar to traditional optical character recognition” — it is the classic approach, and it behaves like one.

The Accurate path uses a neural network to find text as strings and lines first, and only then analyses those for individual words and sentences. Apple notes this is much more in line with how humans read, which is the practical difference: it uses the shape of the line to decide what a smudged character probably is, where the fast path is guessing one glyph at a time.

Either path can then run an optional language-correction phase based on natural language processing, to reduce misreadings.

Despite the names, the choice is not really about speed for a one-off page. It matters when the input is awkward. A clean PDF export or a flatbed scan gives the fast path everything it needs. A photograph taken at an angle, a receipt, a page with a fold through it: that is where line-level context does the work, and the accurate path is what you want.

One consequence worth knowing if you are drawing boxes around the results: the two paths produce different bounding rectangles. The fast path derives them from individual characters; the accurate path tokenises on whitespace, which Apple warns means Chinese text will likely yield lines or fragments rather than complete pieces of text.

The language rules are stricter than they look

This is the part that costs people an afternoon, and it is all in one section of the documentation.

By default, Vision biases its results toward English. If your documents are not in English and you have not said so, you are not getting the accuracy the framework is capable of. You set recognitionLanguages with an array, and the order matters — it sets relative importance, not just membership.

Which languages are available is not a fixed list either. It depends on the path you chose and the API revision, which is why Apple exposes supportedRecognitionLanguages(for:revision:) as a class method to query rather than publishing a table. Call it rather than assuming.

Chinese has its own rules, and they are restrictive. To recognise traditional and simplified Chinese you specify zh-Hant and zh-Hans as the first elements of the array, and — the line that surprises people — English is the only other language you can pair with Chinese. There is no Chinese-plus-Japanese configuration. On top of that, Vision supports neither language correction nor the customWords list for Chinese at all, so the correction step that cleans up the other languages simply is not there.

That customWords property is worth knowing for English work, though. If your documents are full of drug names, part numbers or legal jargon, you can bias the correction phase toward a list of terms you supply. It is ignored unless language correction is enabled, which is an easy thing to get wrong and never notice.

What a model buys you that Vision does not

So when is an OCR model worth installing? The answer is in what MiniCPM-OCR-1.0’s card describes, and it is not “better character recognition”.

It is a roughly 1.3B-parameter model trained for OCR on top of MiniCPM-V 4.6, pairing a SigLIP vision encoder with a Qwen3.5 language model, and it is Apache 2.0. What it does beyond reading text is the point: tables, formulas, charts, seals, and key-information extraction with JSON output, driven by a PP-DocLayoutV3 layout model. Its examples extend to handwritten notes.

That is a different job. Vision hands you strings and the boxes they came from. A document model hands you an invoice as fields, a table as a table, and a formula as something other than a mangled line of symbols. If what you have is a thousand purchase orders and what you want is structured rows, the model is doing work that Vision was never designed to do, and no amount of post-processing on plain strings gets you there cleanly.

On its own benchmarks it scores 93.68 on Real5-OmniDocBench, a real-world document parsing benchmark, and 54.7 on OCRBench v2. Those are the publisher’s figures, and the gap between them is itself informative about how much these scores depend on which benchmark you pick.

What it costs you

Three things, and they are all on the same card.

Chinese and English only. For a model competing with a framework whose language support runs much wider, this is the constraint that decides it for a lot of people before anything else does.

No minimum hardware is published. The card states throughput of about 1.2 pages per second — measured on an NVIDIA RTX 5090. That is not a figure you can carry to a laptop, and the card is explicit that real throughput depends on content, backend, concurrency and hardware. If you are planning to run this on a Mac, the honest position is that nobody has published what it does there.

The card does put its speed in context against other document models, reporting roughly 1.49 times the throughput of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro, with an optional DSpark decoding mode claimed to improve end-to-end efficiency by a further 55%. Those are comparisons against other models doing the same heavyweight job, not against the OCR already on your Mac, and they are the publisher’s own measurements.

It is a project, not a feature. The runtimes listed are vLLM, SGLang, llama.cpp, Ollama and FlagOS, with GGUF, AWQ and GPTQ quantisations to choose between. That is a reasonable evening for someone comfortable with running local models through Ollama, and a wall for everyone else — against a framework that is already installed and needs no choices at all.

If you are going to write the code

Vision is a developer framework rather than an app, which is the one real barrier here, and Apple’s documentation gives a starting point for each path rather than leaving you to assemble one. The fast path has a worked sample in “Extracting phone numbers from text in images” — a good shape to copy because phone numbers are exactly the case where character-level recognition plus a pattern match beats anything cleverer. The accurate path has “Structuring recognized text on a document”, which is the one to read if your output needs to keep any relationship to the layout it came from.

The retrieval pattern in both is the same: you get observations back, ask each for its top candidates, and take the string. The useful habit is to keep more than the first candidate when the stakes are high — the framework is telling you it had alternatives, and discarding them throws away the only signal you get about its confidence.

Choosing between them

Look at the document, not the technology.

If you want the words out of a page — a scan, a photo of a sign, a screenshot, a PDF that will not let you select text — Vision wins on every axis that matters. It is there, it is free, it is on-device, and it handles far more languages.

If you want the structure — the table as rows, the invoice as fields, the formula as notation, the form as key-value pairs in JSON — that is the model’s job, and it is worth the setup if the volume justifies it. One invoice does not. Four hundred a month does.

If your documents are not in English or Chinese and you need structure, you are currently between the two options, and the honest answer is that this particular model is not for you yet.

Where both of them fail

Neither route solves the physical problems. A curved page in a photograph, a fold through a line of text, a receipt printed on thermal paper that has been in a wallet for a month — these degrade both approaches, and the fix is a better capture rather than a better model. Flatten the page, light it evenly, shoot straight on.

Handwriting is the other shared weakness, and it splits the two routes in an instructive way. Apple’s text-recognition page is written around printed and photographed text; MiniCPM-OCR’s card includes handwritten notes among its examples and even a fine-tuning guide for French handwritten text lines. Fine-tuning for a specific hand is a real capability and a real amount of work, and it is worth knowing it exists before you conclude that a shoebox of handwritten receipts is hopeless.

Multi-column layouts remain the classic trap for plain OCR: Vision gives you strings and their positions, and reading order across columns is something you have to reconstruct, which is precisely the gap a layout model fills.

And the failure mode they share is the important one: neither raises a flag when it guessed. Vision at least hands back ranked candidates, which is a signal if you keep them, but take only the top string and a 5 that was really an S arrives looking exactly like a 5 that was really a 5. A model returning well-formed JSON gives you even less to go on, because the structure looks equally correct either way. If the output feeds something that matters — payments, records, anything with a number in it — the checking step is yours to build, in the same way that a decision model’s probabilities are not a substitute for testing on your own data.

Start with the OCR already on the machine. Reach for a model when you can name the structure you need and count the documents you need it from.

Sources

Every figure in this article traces to one of these. Dates are when the page was read.

  1. Recognizing Text in Images
    AppleAccessed Oct 31, 2026Official docs
  2. MiniCPM-OCR-1.0 model card
    OpenBMBAccessed Oct 31, 2026Official docs
  3. Trending models
    Hugging FaceAccessed Oct 30, 2026Official docs

Written by

Kai Devlin AI Automation Engineer

Full-stack developer who builds and runs AI automation systems in production. Runs local LLMs on personal hardware, builds N8N pipelines that actually ship, and deploys on Cloudflare Pages. Every guide on Pipeline Monk is tested on real consumer hardware — a Ryzen 7 5800HS with 16GB RAM and a GTX 1650. If it works on that, it works.