Introducing Surya OCR 2.5! - Most accurate open OCR model under 1B params (85.9 on OmniParseBench) - 1.7x faster than Surya 2, 4.5 pages/s on H100 - 689M params, with character tokenization

You can find Surya here: Model - Code - Blog -

One of the unique things we do with the Surya OCR model is character tokenization. This makes the model smaller and more efficient. With 2.5, we struck a balance by also adding some merged tokens.

Repetitive loops are common in small OCR models, and we've spent a lot of time cleaning them up in training. On a hard set, Surya OCR 2.5 loops 38x less than Surya OCR 2.

Surya OCR 2.5 remains a strong multilingual model. 52 languages score over 80% in our internal benchmark. Over the last couple of years, we've seen Surya be used in low-resource language research. We hope it will continue to be valuable in this context.

Congrats to @hheiden12 on the release! We'll have many more updates to Chandra and Surya soon.

Is there a way to access the dataset to check how it did on different types of files?

I'm trying out bulk processing with Open models, and Chandra, Indra, Surya comes up often: all from your team, and it's a bit confusing for newbies like me... Which one should we use for which kind of tasks?

Yes, good question: - Indra - 1M params, CPU, very fast - Surya - 600M params, good for consumer GPUs - Chandra - 5B, when accuracy is critical or larger GPUs

@SNB_ZZ Any that you would recommend specifically for 19th-century documents? Newspapers, Congressional reports, that sort of thing.

@SNB_ZZ Chandra will likely be best on those

@Presidentlin

Sub-billion OCR still sets the floor

This is really great stuff !!

@vanstriendaniel I loved Surya 2, but wasn’t able to use it for my project in the company due to commerical license issues. Can you please release fully open version.

Congratulations! tested and made a video

Let's give it a try. Thanks!

When compared with Indra OCR?

how is it on messy handwriting

85.9 on OmniParseBench, but that's an aggregate. per-category recall on dense tables? that's where OCR actually falls over.

how much of the 85.9 leans on full-page regeneration? the bench runs it on but the package default is off

官博表格那栏写死了：2.5 只吐 HTML，行列坐标几何没了；要框还得回 2。

Congrats! Just wondering when would there be a newer flagship. Coz for a lot of people, accuracy is much more important than the size of the model.

Congrats on Surya 2.5, Vik. Making document extraction faster and more accessible is valuable work. Keep it going. If you're looking to reach more developers through creators in the space, happy to help

Can it do strike through?

Impressive leap in OCR performance! Lightweight, faster, and more accurate models are exactly what the AI ecosystem needs. 🚀

Surya OCR 2.5：1B参数以下最准确＋1.7倍速度，开源OCR新标杆。这波性能数据，真的很有诚意！