This week, we launched Indra, a 1M parameter OCR model that's 15x faster than Tesseract on CPU and much more accurate. I still can't believe how good this model is.

Check it out here: Code:

When will the Arabic language be available?

Looking into it!

that's a steal give me 2

It's very very good. So good I ported it to WASM and I'm running it in prod:

Can you share how to classify documents into the ones which need indra versus heavier surya/chandra? Can you open source the document classifier, the one you use on alternatively share some insights how to build one

Hey Vik it works great. But on further tests it fails to recognize anything on slightly skewed pages like 7 desgress. On fully rotated pages however it works as its supposed to

We'll look into it, thanks

What’s the Hebrew amount of data you have , and theoretically how much would you need of it more ?

This model doesn't do hebrew, but future versions should. Feel free to DM if you have pointers to data!

Hi vik, it is indeed great! Faster and more accurate than Apple Vision on ANE on my machine. Any plans for additional finetuning on math/formulas/equations?

Yes, we'll look into this

Queue scans overnight on a spare box, cpu ocr loves that

If text is rotated, it fails to read. Imagine a text on Indian High School Certificate - Text in Semi Circle & Caligraphy, the output is not accurate on those text. How can we solve those cases?

It should work on rotations

Which languages does it support? Is that good in Turkish language?

What all scripts does it cover?

This currently covers Latin, Cyrillic, Greek, Devanagari, Bengali, and Chinese.

Does it do piano sheets?

Ollma gwen indra, my first local rag

1M params beating Tesseract on CPU is kind of absurd. how does it handle crappy phone photos of receipts? that's where most OCR falls apart

Does it support Tamil ?

where do you think Indra's headline score to leader model make a difference ?

Amazing! Can't wait to put it through some tests. Does the devanagari recognize Vedic svaras?

Would you be open to a chat?

Feel free to DM!

@grok is it open source?

NT Hao

Is there a blog coming on how it was trained? :)

Is it open source?

Nowadays seeing a n number of ocr models

i'd trust the speed more than the accuracy on messy scans

congrats!

Apache-2.0 license 😍🙏🙏🙏

Do you mean a 1M (post title) param model or 10M (chart number)?

The OCR model is 1.24M params, the tested config in the chart is with a special math model and a few others that are collectively 12M params

Ah my bad, didn’t read the subtext. Very impressive!

1M parameters and 15x faster on CPU is the kind of launch that quietly ends up everywhere. Congrats.

my OCR model is 15x faster than Tesseract. meanwhile my phone's keyboard still thinks "duck" is what i meant. 💀

It's amazing, and to think that 6 years ago I was using 2 giant models to have a single extraction

The weights aren’t free though.