Reads Hindi and Marathi
Gati Vision v1 turns scanned Devanagari pages into plain text. On pages 900px and wider, Hindi and Marathi score the same.
83.4% Hindi, 900px and widerBenchmarks
Tested on scanned Hindi,
Marathi and English pages.
Character error by language and scan widthLower is better · median character error rate %
16.6Gati Vision v1
16.6%
16.6%
29.8Gati Vision v1
29.8%
29.8%
16.9Gati Vision v1
16.9%
16.9%
37.9Gati Vision v1
37.9%
37.9%
1.5Balsa Vision v1
1.5%
1.5%
1.3Gati Vision v1
1.3%
1.3%
5.3Balsa Vision v1
5.3%
5.3%
2Gati Vision v1
2%
2%
Every check, in detail23 checks
| Measure | Balsa Vision v1 | Gati Vision v1 |
|---|---|---|
| Hindi, median character errorTypical share of Hindi characters read wrong | 252.9% | 18.1% |
| Hindi, mean character errorAverage Hindi error, including pages that fail | 299.0% | 31.5% |
| Hindi, 900px and widerTypical Hindi error on scans 900px or wider | Not reported | 16.6% |
| Hindi, under 900pxTypical Hindi error on narrower, lower resolution scans | Not reported | 29.8% |
| Hindi, failed pagesHindi pages with more errors than characters | 20 of 20 | 5 of 103 |
| Hindi, characters in the wrong scriptHindi characters written in another Indian script | No Devanagari output | 2.72% |
| Hindi, mean time per pageAverage seconds to read one Hindi page | 36.7 s | 9.9 s |
| Marathi, median character errorTypical share of Marathi characters read wrong | 226.2% | 32.5% |
| Marathi, mean character errorAverage Marathi error, including pages that fail | 229.3% | 69.8% |
| Marathi, 900px and widerTypical Marathi error on scans 900px or wider | Not reported | 16.9% |
| Marathi, under 900pxTypical Marathi error on narrower, lower resolution scans | Not reported | 37.9% |
| Marathi, failed pagesMarathi pages with more errors than characters | 20 of 20 | 15 of 103 |
| Marathi, characters in the wrong scriptMarathi characters written in another Indian script | 0.00% | 0.15% |
| Marathi, mean time per pageAverage seconds to read one Marathi page | 32.3 s | 10.1 s |
| English, median character errorTypical share of English characters read wrong | 1.6% | 1.4% |
| English, mean character errorAverage English error, including pages that fail | 6.6% | 7.3% |
| English, 900px and widerTypical English error on scans 900px or wider | 1.5% | 1.3% |
| English, under 900pxTypical English error on narrower, lower resolution scans | 5.3% | 2.0% |
| English, pages at 5% error or betterEnglish pages with at most 5 wrong characters in 100 | 84 of 103 | 90 of 103 |
| English, failed pagesEnglish pages with more errors than characters | 1 of 103 | 1 of 103 |
| English, mean time per pageAverage seconds to read one English page | 5.4 s | 6.7 s |
| Peak memoryMost memory used while reading a page | 1.61 GB | 1.72 GB |
| On diskSize of the model files on disk | 631 MB | 715 MB |
The base model is an open document model released under Apache 2.0. Its publisher lists Hindi in Devanagari script among its languages; Marathi is not named. Measured on 309 real scanned pages from printed books, 103 per language, each proofread by a person. Balsa Vision v1 ran on 20 Hindi and 20 Marathi pages and failed every one, so Hindi and Marathi pages go to Gati Vision v1. A failed page has more errors than characters. Character error can go above 100% when a model adds text that is not on the page. How we test
Capabilities
One model,
four jobs done well.
English too
It reads English pages as well, so one model covers all three languages.
98.7% English, 900px and widerTakes the Devanagari pages
Balsa Vision v1 fails on every Hindi and Marathi page tried. Gati Vision v1 reads those pages.
20 of 20 Hindi pages failed by Balsa Vision v1Runs on your machines
A 0.96B model in 4-bit weights. It needs 1.72 GB of memory at peak, so no page leaves the device.
715 MB on diskModel card
715 MB on disk.
On device, no cloud.
Gati Vision
Gati Vision v1 is a document reader for India. It turns scanned Hindi, Marathi and English pages into plain text, on device.
- Built for
- India
- Built on
- An open model (Apache 2.0)
- Base model languages
- Hindi in Devanagari script is listed. Marathi is not named. As published by the base model's publisher.
- Domain
- Document conversion, in Hindi, Marathi and English. Tested on printed book pages.
- Version
- v1, November 2026
Spec
- Model
- gati-vision-1b-v1
- Parameters
- 0.96B
- Weights
- 4-bit
- Memory
- 1.72 GB peak
- Languages
- Hindi, Marathi, English
- Scripts
- Devanagari and Latin
- Output
- Plain page text
- Cloud
- Not needed
- Runs on
- Apple silicon, on device
Mean time per page
9.9s
About 363 Hindi pages an hour on a 16 GB laptop.
Repeatability, Hindi
10/10
Re-run Hindi pages gave the same text, byte for byte. All 31 re-runs across three languages matched.
Built for India
- Hindi accuracy, pages 900px and wider83.4%
- Marathi accuracy, pages 900px and wider83.1%
- Peak memory1.72 GB
- Same text on re-run31 of 31
Film
Watch it run
on device.
How it reads.
- 01Look at the page
- 02Find the text
- 03Write it out
- 04Keep the layout
Playground
Try it yourself.
Watch it work, step by step.
Pick a step or let it play. Demos tagged as illustrations show example output, and the scores on this page come from our own tests.
More Gati models
Five models,
each doing one job.
Gati Omni
Gati Omni v1 is a multimodal model for India, with 4.5B effective parameters. It reads text, images and audio, replies in eight languages used in India and runs on device.
4.5BView modelGati TextGati Text v1 is a small assistant model for India. It answers in Hindi, Marathi, Gujarati, Telugu, English and Hinglish, on device, with no cloud.
2.3BView modelGati VisionGati Vision v1 turns scanned Hindi, Marathi and English pages into text for India, on device with no cloud. It reads best on sharp scans 900px and wider.
1BYou are hereGati VoiceGati Voice v1 reads UPI payment messages aloud in Hindi, Marathi, Hinglish and Indian English, on device, faster than speech on a MacBook Air M5.
2.5BView modelGati DecisionGati Decision v1 reads Hindi payment messages for India and recommends the first step, on device with no cloud. A person makes the final decision.
0.6BView modelQuestions
Asked often.
Gati Vision v1 is a 0.96B document reader for India. It turns scanned Hindi, Marathi and English pages into plain text.
No. Gati Vision v1 runs on device, so no page leaves the machine. Peak memory is 1.72 GB.
Gati Vision v1 ran on a MacBook Air M4 with 16 GB of memory. It reads a Hindi page in 9.9 s on average and takes 715 MB on disk.
Gati Vision v1 reads Hindi, Marathi and English, in Devanagari and Latin script. Balsa Vision v1 fails on Hindi and Marathi pages, so Gati Vision v1 reads them.
Gati Vision v1 was tested on 309 real scanned pages, 103 each in Hindi, Marathi and English, each proofread by a person. On pages 900px and wider it reads 83.4% of Hindi characters and 83.1% of Marathi characters right.
Write to research@f-r.co.
