Point your phone at a page, a notice or a screenshot, and Loa turns the Dhivehi into text you can copy, search and translate. It is Akuru's own model, trained from nothing in the Maldives. Loa (ލޯ) means eye.
This is Loa's real reading of the sample below, step by step.
First words, step by step: repeats and empty steps fade out
ރާއްޖޭގެ ރަށްރަށުގައި ދިރިއުޅޭ ރައްޔިތުންނަށް
Step 1. Loa straightens the page and finds each line of text, keeping the fili above and below with their letters.
Four images drawn in different fonts and styles, read by Loa and by Tesseract, the best-known open OCR with a Dhivehi model. The readings are real and unedited.
Faruma · 15px
އަކުރު AI އަކީ ދިވެހި ބަހަށް ހާއްސަކޮށް ތައްޔާރުކުރި ޒަމާނީ ޓެކްނޮލޮޖީއެކެވެ.ފޯނުގެ ސްކްރީނުން ނަގާ ކުދި ލިޔުންތައް ވެސް ކިޔައިދެއެވެ.
އަކ ދިވެ ހ ބަހަށް ހާއްސަކޮ ށް ތައްޔާރުކު ރ ޒަމާ ނ ޓެކްނޮލޮޒީއެކެވެ.އަކުރު 1 އަކ ދވެ ހ ަހް ހ ރި ޒަމާނިފޯނުގެ ސްކްރިނު ނާ ނަޤާ ކުދި ލޔުނާތައް ވެސް ކިޔައިދެއެވެ.ފިޔަގ ީ ޙީީ
Mv GroupX Avas · 30pxރާއްޖޭގެ ރަށްރަށުގައި ދިރިއުޅޭ ރައްޔިތުންނަށްޚިދުމަތް ފޯރުކޮށްދިނުމަކީ މުހިންމު ކަމެކެވެ.
ޘ ބބ ގ އ ބ ލޢ މބ ވ ބ މ ބި އ ބޗއ ޒ އ ޗނތނ މފ ދއމ ތއތނނތމ ވ ބި ބމ »ބ މތ ބ މ ހ ގ ގ >ހދމާޗ ރތކނދިނހުތި މނމ ކމކ ވٌ
MV Bodu · 54pxދިވެހި ބަސް ދިރުވައި ހިފަހައްޓަމާ
¬ ީބ ސ ރ ޢިއ ހިބީދިވެހި ބސޮ ދިރަވައި ހިފިހއެޓޫމާ
ak-rasmee · 26px
2026 ވަނަ އަހަރުގެ އޮކްޓޫބަރު 28 ގައި Loa v2 ނެރުނީކުށުގެ މިންވަރު 0.31% އަށް ދަށްވިއެވެ.
ހިހި ` ިޚި މިޒި ިބިރހި މި ހ ބމ6 ަނ އަހަރުގެ އޮކްޓޫބަ ރު 8 ގައި ?† 164 ނާރުނިމިމިބި » ހިމި އި ީބ« _ ބިބީކުށުގެ މަންވަ ރު 031 އަ ޔާ ދަންވިއެވެ.
Tesseract is the most widely used open OCR engine, and its official Dhivehi model is what most Dhivehi OCR runs on today. Both read the same pages.
| 11px | 13px | 16px | 20px | 28px | 40px | 64px | 96px | |
|---|---|---|---|---|---|---|---|---|
| Loa · Screen | 0.2% | 0.1% | 0.4% | 0.0% | 0.4% | 0.6% | 0.0% | 0.8% |
| Loa · Scan | 1.1% | 0.1% | 0.4% | 0.0% | 0.7% | 0.0% | 0.3% | 0.0% |
| Tesseract · Screen | 58.1% | 34.4% | 16.7% | 30.3% | 100% | 93.8% | 100% | 100% |
| Tesseract · Scan | 51.6% | 39.1% | 19.1% | 57.6% | 100% | 95.8% | 100% | 100% |
A new line reader, starting from random weights, trained on about a million lines of Dhivehi drawn by our own data engine. Near word-perfect on Dhivehi body text; English and numbers still failed.
200,000 new lines with dates, times and reference numbers. First score on the new 32-page exam.
1.5 million new lines: 20 Thaana fonts, English and digits in Latin fonts, tiny phone-screen text and large titles, all in the order the Unicode standard draws them. Now in the app.
More headline fonts, whole book and newspaper layouts, and handwritten Thaana.
We draw Dhivehi sentences from public text in 20 fonts and many sizes, then age them like real images: blur, noise, tilt, paper tone. Every line comes with its exact answer, so Loa learns from millions of perfectly labelled examples.
Loa is a 14-million-parameter network: a convolutional eye that looks at the line, and a two-way memory that reads it in context. It reads a page in a few seconds, even without a GPU.
Every version sits the same fixed exam before it ships. A new version goes live only if it beats the one in the app.
Loa runs on Akuru's own servers. Your images are not sent to other AI companies.
Ministries, banks, newsrooms, schools and app makers: read Dhivehi documents at scale with one API call. Self-serve keys, the same credits as the app, free credits every month to start.
curl https://api.akuru-ai.com/v1/ocr -H "Authorization: Bearer $AKURU_API_KEY" -d '{"image": "'$(base64 -w0 page.jpg)'"}'{
"model": "akuru-loa-T-2.0",
"text": "ދިވެހި ބަސް ދިރުވައި ހިފަހައްޓަމާ"
}Loa is Akuru's optical character recognition (OCR) model for Dhivehi. It reads printed Thaana in photos, scans and screenshots and turns it into editable Unicode text.
Open the Akuru app, tap +, and add a picture. Loa reads it and you can copy, translate or listen to the text. New accounts get free credits every month.
It was trained on 20 Thaana fonts, including Faruma, MV Waheed, MV Bodu, AK Rasmee, Avas and Reethi, from 9 px screen text to large titles, plus English and numbers in the same line.
Not yet. Loa is built for printed and on-screen text. Handwriting is on the plan.
No. Loa was trained from scratch by Akuru. It does not use another company's model or service.
Take a photo of any Dhivehi page and see Loa at work.