What is needed to evaluate OCR for most business applications (above everything else) is accuracy.
Some results look plausible but are just plain wrong. That is worse than useless.
Example: the "Table" sample document contains chemical substances and their properties. How many numbers did the LLM output and associate correctly? That is all that matters. There is no "preference" aspect that is relevant until the data is correct. Nicely formatted incorrect data is still incorrect.
I reviewed the output from Qwen3-VL-8B on this document. It mixes up the rows, resulting in many values associated with the wrong substance. I presume using its output for any real purpose would be incredibly dangerous. This model should not be used for such a purpose. There is no winning aspect to it. Does another model produce worse results? Then both models should be avoided at all costs.
Are there models available that are accurate enough for this purpose? I don't know. It is very time consuming to evaluate. This particular table seems pretty legible. A real production grade OCR solution should probably need a 100% score on this example before it can be adopted. The output of such a table is not something humans are good at reviewing. It is difficult to spot errors. It either needs to be entirely correct, or the OCR has failed completely.
I am confident we'll reach a point where a mix of traditional OCR and LLM models can produce correct and usable output. I would welcome a benchmark where (objective) correctness is rated separately from of the (subjective) output structure.
Edit: Just checked a few other models for errors on this example.
* GPT 5.1 is confused by the column labelled "C4" and mismatches the last 4 columns entirely. And almost all of the numbers in the last column are wrong.
* olmOCR 2 omits the single value in column "C4" from the table.
* Gemini 3 produces "1.001E-04" instead of "1.001E-11" as viscosity at T_max for Argon. Off by 7 orders of magnitude! There is zero ambiguity in the original table. On the second try it got it right. Which is interesting! I want to see this in a benchmark!
There might be more errors! I don't know, I'd like to see them!
I didn't expect IBM to be making relevant AI models but this thing is priced at $1 per 4,000,000 output tokens... I'm using it to transcribe handwritten input text and it works very well and super fast.
I suggest you make explicit the assumption that this website is specifically about English text. Otherwise the leaderboard is pretty meaningless, with extreme differences in performance across other scripts - and potentially even languages such as Vietnamese or Czech which use Latin but have lots of accents.
Hey! I'm the dev who made this:) I think that you are right, data will bias towards english because we have a dataset that people can use that is in english. But you can also upload non-english docs into the battle mode as well as the playground!
LMArena splits their leaderboard by language: maybe you should consider doing the same thing
I assume to do that you’d need another model to do language detection on the inputs and/or outputs; but a language detection model can be a lot cheaper than an OCR model or an LLM
That's unfortunate because I have a bunch of photos with handwritten German on the back that I need to transcribe, and seeing as that I can't read German I can't really do it by myself either.
I reckon performance on German will be similar to English, the only real difference is the umlauts and those are very consistent. Not sure how it will do on the ß.
i don't think i'm you're target audience but i found it interesting to see the side-by-side comparisons from images with text in. it's pretty cool to see how different models interpret photos, too. cool tool, must've been fun to make.
I'm very impressed by the models, to the point I was wondering if they were really converting the pdf or just reading the content. I tried on documents in french, english and spanish, very heaving on graphics and with complex layouts (boardgame, flyer, book about rust), and I wasn't expecting anything great. Especially some models were showing symbols and smileys quite close from the original.
I noticed that some models were resisting better to faking data than other, especially I saw that in a sentence cut from the document, GPT5 was inventing the end of the sentence and opus was properly showing it cut.
I didn't try with my writing but in the playground there is one example and some models read it better than me.
I wish the output would show the confidence of the model on each part. I think it would help immensely.
Note that sometimes a model get stuck in a loop, preventing to vote and to see which model is which
There have been such a large number of OCR tools pop up over the past ~year; sorely in need for some benchmarks to compare them. Would love to see support for normal OCR tools like tesseract, EasyOCR, Microsoft Azure, etc. I'm using these for some projects, and my experiments with VLMs for OCR have resulted in too much hallucination for me to switch. Benchmarks comparing across this aisle would be incredibly useful.
A limitation of this leaderboard approach that I want to point out is that while the large general-purpose LLMs can make greater leaps of inference (on handwriting and poor quality scans), and almost always produce better layouts and more coherent output, they can also sometimes be less correct. My experience is that they're more prone to skipping or transposing sections of text, or even hallucinating completely incorrect output, than the purpose-trained models. (A similar comparison can be made in turn to the character- or word-based OCR approaches like Tesseract, which are even less "intelligent" but also even less prone to those malbehaviors.)
Also, some of the models are prone to infinite loops and I suspect this is not being punished appropriately; the frontend seems to get into a bad state after around 50k characters, which prevents the user from selecting a winner. Probably would be beneficial to make sure every model has an output length limit.
Still, a really cool resource - I'm looking forward to more models being added.
Totally agree w/ your first point! For the looping, we just added a stop condition for now in battle mode, and you can still vote on the other model afterwards. A bit of a hard problem to solve. We will add more models!
If the text is written interactively on the canvas (as opposed to extraction from pixels) this task is known as "online handwriting recognition" ("online" because you can watch the text being formed incrementally, which makes it easier to e.g. distinguish individual strokes.)
I don't know what the state of the art is, but an old model for digitizer pens might not do so bad either.
Love this! Would have liked to see something like textract for a pre-LLM benchmark (but of course that's expensive), and also a distinction between handwritten text and printed one.
Interesting that the 8B of the Qwen3-VL family 9th place, above a few proprietary models. This thing can run locally with llama.cpp on modest hardware.
UX on mobile isn’t great. It wasn’t obvious to me where the second model output was and I was thrown off even more so because the option to vote for model 1 output was presented without ever even seeing model two output.
Second suggestion would be to install a MathJax plugin so one can properly rate mathematical equations and formulas. Raw LATeX is easy to mistake and it makes comparing between LATeX and Unicode outputs hard.
Really like the idea. Unfortunately, my first upload is still spinning on one of the models about 5 minutes in. Clicking "Stop Battle" seems to do nothing either
We wanted to keep the focus on (1) foundation VLMs and (2) open source OCR models.
We had Mistral previously but had to remove it because their hosted API for OCR was super unstable and returned a lot of garbage results unfortunately.
Paddle, Nanonets, and Chandra being added shortly!
MistralOCR works stably for me when first uploading the file to their server and then running the OCR. I also had some issues before when giving a URL directly to the OCR API, not sure if you're doing that?
Ultimately, there’s some intersection of accuracy x cost x speed that’s ideal, which can be different per use case. We’ll surface all of those metrics shortly so that you can pick the best model for the job along those axes.
Some results look plausible but are just plain wrong. That is worse than useless.
Example: the "Table" sample document contains chemical substances and their properties. How many numbers did the LLM output and associate correctly? That is all that matters. There is no "preference" aspect that is relevant until the data is correct. Nicely formatted incorrect data is still incorrect.
I reviewed the output from Qwen3-VL-8B on this document. It mixes up the rows, resulting in many values associated with the wrong substance. I presume using its output for any real purpose would be incredibly dangerous. This model should not be used for such a purpose. There is no winning aspect to it. Does another model produce worse results? Then both models should be avoided at all costs.
Are there models available that are accurate enough for this purpose? I don't know. It is very time consuming to evaluate. This particular table seems pretty legible. A real production grade OCR solution should probably need a 100% score on this example before it can be adopted. The output of such a table is not something humans are good at reviewing. It is difficult to spot errors. It either needs to be entirely correct, or the OCR has failed completely.
I am confident we'll reach a point where a mix of traditional OCR and LLM models can produce correct and usable output. I would welcome a benchmark where (objective) correctness is rated separately from of the (subjective) output structure.
Edit: Just checked a few other models for errors on this example.
* GPT 5.1 is confused by the column labelled "C4" and mismatches the last 4 columns entirely. And almost all of the numbers in the last column are wrong.
* olmOCR 2 omits the single value in column "C4" from the table.
* Gemini 3 produces "1.001E-04" instead of "1.001E-11" as viscosity at T_max for Argon. Off by 7 orders of magnitude! There is zero ambiguity in the original table. On the second try it got it right. Which is interesting! I want to see this in a benchmark!
There might be more errors! I don't know, I'd like to see them!
I didn't expect IBM to be making relevant AI models but this thing is priced at $1 per 4,000,000 output tokens... I'm using it to transcribe handwritten input text and it works very well and super fast.
Super nice if it worked for our use case to simply get full output.
I assume to do that you’d need another model to do language detection on the inputs and/or outputs; but a language detection model can be a lot cheaper than an OCR model or an LLM
i have it verify some stamps which are quite messy and sometimes obscured and honestly some i could not even read.
I noticed that some models were resisting better to faking data than other, especially I saw that in a sentence cut from the document, GPT5 was inventing the end of the sentence and opus was properly showing it cut.
I didn't try with my writing but in the playground there is one example and some models read it better than me.
I wish the output would show the confidence of the model on each part. I think it would help immensely.
Note that sometimes a model get stuck in a loop, preventing to vote and to see which model is which
Also, some of the models are prone to infinite loops and I suspect this is not being punished appropriately; the frontend seems to get into a bad state after around 50k characters, which prevents the user from selecting a winner. Probably would be beneficial to make sure every model has an output length limit.
Still, a really cool resource - I'm looking forward to more models being added.
Working on a hobby project that interacts with user handwriting on <canvas>. Tried some CNN models for digits but had trouble with characters.
Note that I haven't tried any of them, but tesseract is still likely the leading open source OCR that works with CPU.
I don't know what the state of the art is, but an old model for digitizer pens might not do so bad either.
But still, this is incredibly useful!
Just this morning I came across HunyuanOCR which sounded very promising. https://huggingface.co/tencent/HunyuanOCR
UX on mobile isn’t great. It wasn’t obvious to me where the second model output was and I was thrown off even more so because the option to vote for model 1 output was presented without ever even seeing model two output.
Second suggestion would be to install a MathJax plugin so one can properly rate mathematical equations and formulas. Raw LATeX is easy to mistake and it makes comparing between LATeX and Unicode outputs hard.
I’ve had great results locally. Albeit you need macOS >=13 for this.
We had Mistral previously but had to remove it because their hosted API for OCR was super unstable and returned a lot of garbage results unfortunately.
Paddle, Nanonets, and Chandra being added shortly!
[see https://news.ycombinator.com/item?id=45988611 for explanation]