North MicroVision OCR Accuracy: What Real Tests Show
Hands-on testing of Cohere's North MicroVision shows strong invoice and French OCR, shaky handwriting results, and weak Urdu and Indonesian accuracy.
How accurate is North MicroVision at OCR?
North MicroVision, an open weight vision language model from Cohere, performs unevenly depending on what you feed it. In hands-on testing across handwritten notes, physics equations, a printed invoice, French cursive, and Arabic newspaper text, the model nailed clean printed English and European-language documents but struggled badly with messy handwriting and non-European scripts, particularly Urdu and Indonesian. It’s a 2.4 billion parameter model, so expectations should be calibrated accordingly, but the gap between its strongest and weakest results is wide enough to matter for anyone planning to deploy it.
TL;DR
- Printed structured documents are where it shines: on a printed invoice with a table, the model extracted every field, address, and line item with the totals matching exactly, a flawless result.
- Handwriting recognition is inconsistent: on a multilingual handwritten note, it missed capitalization cues, added punctuation that wasn’t there, and made outright translation errors in some lines.
- European languages hold up well, in line with Cohere’s broader track record, while Urdu and Indonesian accuracy was poor, with words like “mau” (want) misread as “may” and “roti” confused with “kopi” (coffee).
- Dense Arabic script broke the model in this test, with the transcription falling apart on a busy right-to-left newspaper banner.
- French cursive on a doctor’s letterhead came out nearly perfect, including separating handwritten text from the printed letterhead and stamp, with only minor slips near the end.
- Handwritten physics equations were a mixed bag: LaTeX output for equations like the Lorentz factor and E=mc² was clean and correct, but the model lost coherence and repeated itself toward the end of a longer, symbol-dense passage.
- VRAM usage was modest, consuming just over 5 GB when fully loaded, making it practical to run on a single consumer or workstation GPU.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
What is North MicroVision and how does it work?
North MicroVision is an open weight vision language model released by Cohere under the Apache 2.0 license, meaning it can be used, modified, and redistributed freely, including for commercial purposes. It packs its capabilities into 2.4 billion parameters, small enough to run locally on a single GPU and, according to Cohere, fine-tune on custom data.
The architecture runs in three stages. First, a native-resolution vision encoder processes an image at its actual size rather than downscaling it, which helps preserve fine detail in documents, charts, and screenshots. Second, a multimodal projector converts those visual features into a representation the language model can reason over. Third, a 2 billion parameter language model reads the combined text and image tokens and generates a response. That language model uses a mix of sliding window attention layers and full attention layers repeated across 28 layers, a design meant to balance inference speed with the ability to track longer-range context.
The “native resolution” framing is the model’s core pitch: rather than treating an OCR problem as a scaled-down thumbnail, it aims to read documents closer to how a person would view the original file.
How does it perform on printed versus handwritten text?
The clearest pattern across testing is that print beats handwriting, sometimes by a wide margin. On a printed invoice with a structured table, the model extracted every field correctly, matched all totals, and produced clean output with zero errors. That’s the kind of task compact vision models are increasingly used for in practice: pulling structured data out of receipts, forms, and invoices.
Handwriting told a different story. On a multilingual handwritten note mixing English, Urdu, and Indonesian, the model missed basic formatting cues like all-caps words, inserted a question mark where none existed in the source, and made several word-level errors even in the English portions. For a 2.4 billion parameter model, the print-versus-handwriting gap suggests it’s most reliable when documents are typed or printed rather than freehand.
Handwritten physics equations produced a more nuanced result. The model correctly rendered several genuinely hard equations in LaTeX, including the space-time interval, the Lorentz factor, and blackbody radiation integrals. But partway through a longer sequence of symbols it lost track and began repeating output, a common failure mode in smaller models asked to sustain structured output over long, symbol-dense sequences.
Which languages does it handle well, and which does it struggle with?
Testing showed a sharp split along language lines. French cursive handwriting on a doctor’s letterhead came out close to perfect, with the model correctly transcribing the cursive text, translating it, and even separating the handwritten note from the printed letterhead and an embedded stamp. Only a few cursive words near the end and some faint stamp detail tripped it up. That result tracks with Cohere’s broader reputation for strong European-language support in its models.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Urdu and Indonesian were a different story. In the handwritten multilingual test, Urdu text wasn’t understood at all. Indonesian fared only slightly better but still contained clear mistranslations: “mau” (Indonesian for “want”) came out as “may,” and “kopi” (coffee) was rendered as “roti” (bread) despite the source clearly showing a coffee reference. A separate test on a dense, right-to-left Arabic newspaper banner produced results that essentially fell apart, failing to deliver a usable transcription or translation.
This unevenness matters for anyone considering the model for multilingual document pipelines. The marketing claim of broad multilingual support doesn’t hold up equally across languages, at least not at this parameter count and not without fine-tuning.
Is North MicroVision worth running locally?
For teams working with printed, structured documents in English or major European languages, the answer looks favorable. The model is small enough to run on a single GPU, reportedly using just over 5 GB of VRAM when loaded, which puts it within reach of a wide range of workstation and even some consumer hardware. Being Apache 2.0 licensed also means no restrictions on commercial use or redistribution, and Cohere has made it available ungated on Hugging Face, unlike some of its other models.
The catch is scope. If the use case involves handwriting, especially in Urdu, Indonesian, Arabic, or other non-European scripts, the current results suggest real limitations. Cohere’s own benchmarks reportedly show the model trading favorably against similarly sized rivals on general visual question answering, chart reading, and document understanding tasks, and the invoice and French cursive results back that up for print-heavy, Western-language workloads. But the newspaper Arabic test and the Urdu/Indonesian handwriting failures suggest the “multilingual” claim needs qualification: strong for some languages, unreliable for others.
Given the model is small and open, fine-tuning on domain-specific or language-specific data is a realistic path to closing some of these gaps, and Cohere has signaled the model is designed to be fine-tunable. For now, though, out-of-the-box accuracy should be treated as language-dependent rather than uniformly multilingual.
Frequently Asked Questions
What is North MicroVision used for?
It’s designed for document understanding tasks: reading forms, invoices, tables, charts, and screenshots, along with general visual question answering. Its native-resolution image processing is aimed at improving accuracy on documents where fine text detail matters.
How many parameters does North MicroVision have?
It has 2.4 billion parameters total, combining a native-resolution vision encoder, a multimodal projector, and a 2 billion parameter language model component.
Can North MicroVision run on a single GPU?
Yes. In testing on an Nvidia RTX A6000, the fully loaded model consumed just over 5 GB of VRAM, well within reach of most modern GPUs used for local inference.
Does North MicroVision support Arabic and Urdu well?
Testing found significant weaknesses here. Urdu handwriting wasn’t understood at all in one test, and a dense Arabic newspaper banner produced a transcription that fell apart. European languages like French performed far better by comparison.
Is North MicroVision open source?
Yes. It’s released under the Apache 2.0 license, allowing free use, modification, and commercial deployment, and the model weights are available without gating on Hugging Face.