Reads English pages
Balsa Vision v1 turns a scanned English page into text. On pages 900px and wider it reads 98.5% of characters right.
98.5% at 900px and widerBenchmarks
Tested on scanned
English book pages.
English character error by scan widthLower is better · median character error rate %
1.4Gati Vision v1
1.4%
1.4%
1.6Balsa Vision v1
1.6%
1.6%
1.3Gati Vision v1
1.3%
1.3%
1.5Balsa Vision v1
1.5%
1.5%
2Gati Vision v1
2%
2%
5.3Balsa Vision v1
5.3%
5.3%
Every check, in detail16 checks
| Measure | Gati Vision v1 | Balsa Vision v1 |
|---|---|---|
| English, median character errorTypical share of English characters read wrong | 1.4% | 1.6% |
| English, mean character errorAverage English error, including the few failed pages | 7.3% | 6.6% |
| English, 900px and widerTypical English error on sharp scans, 900px or wider | 1.3% | 1.5% |
| English, under 900pxTypical English error on narrower, lower resolution scans | 2.0% | 5.3% |
| English, pages at 2% error or betterEnglish pages with at most 2 wrong characters in 100 | 64 of 103 | 61 of 103 |
| English, pages at 5% error or betterEnglish pages with at most 5 wrong characters in 100 | 90 of 103 | 84 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 | 6.7 s | 5.4 s |
| Hindi, median character errorTypical share of Hindi characters read wrong | 18.1% | 252.9% |
| Hindi, failed pagesHindi pages with more errors than characters | 5 of 103 | 20 of 20 |
| Hindi, mean time per pageAverage seconds to read one Hindi page | 9.9 s | 36.7 s |
| Marathi, median character errorTypical share of Marathi characters read wrong | 32.5% | 226.2% |
| Marathi, failed pagesMarathi pages with more errors than characters | 15 of 103 | 20 of 20 |
| Marathi, mean time per pageAverage seconds to read one Marathi page | 10.1 s | 32.3 s |
| Peak memoryMost memory used while reading an English page | 1.68 GB | 1.61 GB |
| On diskSize of the model files on disk | 715 MB | 631 MB |
The base model is an open document model released under Apache 2.0. Balsa Vision v1 is our own conversion of it. Against the reference conversion it scores the same on every one of the 103 pages, and reads them faster, 5.4 s a page against 6.0 s. Measured on 103 real scanned English pages from printed books, each proofread by a person, plus 40 Hindi and Marathi pages for Balsa Vision v1. Same pages and same scorer for both models. A failed page has more errors than characters. Balsa Vision v1 reads Latin script, so Gati Vision v1 reads Hindi and Marathi pages. How we test
Capabilities
One model,
four jobs done well.
Keeps the layout
Tags every block with its place on the page, so paragraphs, tables, code and formulas come out apart. This is the base model's DocTags format.
DocTags output, not scored by usFast on a laptop
It reads about 670 English pages an hour on a 16 GB laptop.
5.4 s a pageRuns on your machines
A 0.26B model in 16-bit weights. It needs 1.61 GB of memory at peak, so no page leaves the device.
631 MB on diskModel card
631 MB on disk.
On device, no cloud.
Balsa Vision
Balsa Vision v1 is a small document reader for the UK. It turns scanned English pages into text with layout tags, on device.
- Built for
- The UK
- Built on
- An open model (Apache 2.0)
- Base model languages
- English, with experimental Japanese, Arabic and Chinese, as published by its maker
- Domain
- Document conversion, in English. Tested on printed book pages.
- Version
- v1, November 2026
Spec
- Model
- balsa-vision-0.3b-v1
- Parameters
- 0.26B
- Weights
- 16-bit, not quantised
- Memory
- 1.61 GB peak
- Language
- English, Latin script
- Input
- Scanned page image
- Output
- Text with layout and position tags
- Cloud
- Not needed
- Runs on
- Apple silicon, on device
Mean time per English page
5.4s
About 670 English pages an hour on a 16 GB laptop, against about 537 for Gati Vision v1.
Repeatability
11/11
Re-run pages gave the same text, byte for byte.
Built for English documents
- Character accuracy, pages 900px and wider98.5%
- Time per English page5.4 s
- Peak memory1.61 GB
- Same text on re-run11 of 11
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 Balsa models
Five models,
each doing one job.
Balsa Omni
Balsa Omni v1 is a multimodal model for the UK, with 4.5B effective parameters. It reads text, images and audio, replies in British English and runs on device, with no cloud.
4.5BView modelBalsa TextBalsa Text v1 is a small assistant model for the UK. It answers in British English and runs on device, on a laptop, with no cloud.
2.3BView modelBalsa VisionBalsa Vision v1 turns scanned English pages into text for the UK, on device with no cloud. It reads 98.5% of characters right on pages 900px and wider.
0.3BYou are hereBalsa VoiceBalsa Voice v1 reads UK customer service messages aloud in British English, on device with no cloud, 21.6 times faster than real time on a MacBook Air M5.
0.04BView modelBalsa DecisionBalsa Decision v1 flags deceived payments and routes UK customer messages, on device with no cloud. It recommends an option, and a person decides.
0.3BView modelQuestions
Asked often.
Balsa Vision v1 is a 0.26B document reader for the UK. It turns a scanned English page into text, with a layout tag on every paragraph, table, code block and formula.
No. Balsa Vision v1 runs on device, so no page leaves the machine. Peak memory is 1.61 GB.
Balsa Vision v1 ran on a MacBook Air M4 with 16 GB of memory. It reads an English page in 5.4 s on average and takes 631 MB on disk.
Balsa Vision v1 reads English in Latin script. It fails on all 40 Hindi and Marathi pages tried, so Gati Vision v1 reads those pages.
Balsa Vision v1 was tested on 103 real scanned English pages, each proofread by a person. It reads 98.5% of characters right on pages 900px and wider, and 94.7% on narrower scans.
Write to research@f-r.co.
