hckrnws

My current holy grail is my attempt to convert a Shipibo (an indigenous Peruvian language)-to-Spanish dictionary into a Shipibo-to-English dictionary. The pdf I have (available freely on archive.org) isn't a great scan (though I think it'd be a heck of a lot easier than some of the handwritten examples they show). Layout (2-columns) along with header/footers can cause some headaches, but it is all Latin script. This seems to fall on its face pretty badly (not even a couple of pages in), so my search continues. (The other major problem I'm having is trying to separate out Shipibo definitions/examples from the Spanish ones, and only translating the Spanish to English...so pretty complex I guess. I've been taking fresh stabs at this project every few months when I see OCR/LLM news pop up and continue to be disappointed)

I'm assuming you're interested in studying Ayahuasca traditions?

I recently learned that traditionally in Shipibo culture, ayahuasca was never meant to be given to "the normal mind". Instead the maestras would be the ones taking the ayahuasca in order to help guide them into diagnosing people dealing with various sicknesses.

These maestras were also ranked by how many different plants they'd done a dieta on. A dieta is kinda similar to fasting. You can't shower with soap, you can't have sex, you can't have too much salt/seasoning, can't be exposed to too much smoke, can't have alcohol, etc. And you use that specific plant throughout your time. Basically you want to eliminate any conflicting variables so you can experience the plant as purely as possible to understand its effects. Traditionally these dietas could last over a year but modern day maestros typically do them for just a few weeks.

I don't really have a point to this. Just found it fascinating how deeply and strictly they study certain plant medicines and wanted to share

Yes essentially. I've got a few resources cobbled together over the last few years but it'd be really nice to have this reference (my Spanish isn't the best, and running to the translator for a definition can be a little annoying). Also to share with fellow learners/apprentices I know. There are a couple of classes out there (which are actually geared more toward the ceremonial/icaro language, not purely conversational Shipibo, which is a bit simpler as you don't need to worry as much about conjugation and other complexities) which I might look into eventually.

(Fwiw I've accumulated a couple years worth of dieta under my belt and am well aware of the restrictions! It's indeed very fascinating, been pretty serious about it the last few years and I've barely scratched the surface)

Couldn't you use your smartphone and Google Lens (on Android, Google app on iOS includes Google Lens functionality) to translate the Spanish to English?

FYI - Lens on Android does in-place language translation including attempting to use the same/similar font that the original language is written/printed.

Unfortunately, I don't think Lens can be used in an automated batch translation mode to convert an entire book/multiple pages

Does it handle math expressions (those rendered from LaTeX) well? I've been looking for a good OCR model to transcribe my math textbooks into markdown (obviously ignoring the images and figures) with LaTeX as math expressions, and none of the current OCR models work reliably enough.

EDIT: you can try it yourself for free at https://console.mistral.ai/build/document-ai/ocr-playground once you create a developer account! Fingers crossed to see how well it works for my use case.

I've just finished processing thousands of documents using the Gemini Pro 3 vision model and it outperformed every OCR and image model I've tested by a long shot, perfect markdown with latex for the math every time.

what prompt are you using?

Please post an update on how well it works for you.

From a tweet: https://x.com/i/status/2001821298109120856

> can someone help folks at Mistral find more weak baselines to add here? since they can't stomach comparing with SoTA....

> (in case y'all wanna fix it: Chandra, dots.ocr, olmOCR, MinerU, Monkey OCR, and PaddleOCR are a good start)

I've worked on document extraction a lot and while the tweet is too flippant for my taste, it's not wrong. Mistral is comparing itself to non-VLM computer vision services. While not necessarily what everyone needs, they are a very different beasts compared to VLM based extraction because it gives you precise bounding boxes, usually at the cost of larger "document understanding".

Its failure mode are also vastly different. VLM-based extraction can misread entire sentences or miss entire paragraphs. Sonnet 3 had that issue. Computer vision models instead will make in-word typos.

Why not use both? I just built a pipeline for document data extraction that uses PaddleOCR, then Gemini 3 to check + fix errors. It gets close to 99.9% on extraction from financial statements finally on par with humans.

I did the opposite. Tesseract to get bboxes, words, and chars and then mistral on the clips with some reasonable reflow to preserve geometry. Paddle wasn’t working on my local machine (until I found RapidOCR). Surya was also very good but because you can’t really tweak any knobs, when it failed it just kinda failed. But Surya > Rapid w/ Paddle > DocTr > Tesseract while the latter gave me the most granularity when I needed it.

Edit: Gemini 2.0 was good enough for VLM cleanup, and now 2.5 or above with structured output make reconstruction even easier.

This is The Way. Remember AI doesn't have to replace existing solutions but can tactfully supplement it.

after clicking on your link I browsed twitter for a minute and damn that place has become weird (or maybe it always was?)

As someone who has been on Twitter since 2007, it’s radically changed in the last few years to the point of being unrecognizable.

I'd want to see a comparison with Qwen 3 VL 235B-A22B, which is IME significantly better than MinerU.

I'm reading worse performance than many OSS offerings like Paddle, MinerU, MonkeyOCR, etc:

https://www.codesota.com/ocr

Thanks for sharing this site

It seems like Mistral is just chasing around sort of "the fringes" of what could be useful AI features. Are they just getting out-classed by OAI, Google, Anthropic?

It seems like EU in general should be heavily invested in Mistral's development, but it doesn't seem like they are.

Form processing is vastly more useful than meme generation. When people need to do real work this is the sort of tool they are going to reach for.

Yep. I saw the title and got excited.... this is a particular problem area where I think these things can be very effective. There are so many data entry class tasks which don't require huge knowledge or judgement... just clear parsing and putting that into a more machine digestible form.

I don't know... feels like this sort of area, while not nearly so sexy as video production or coding or (etc.)... but seems like reaching a better-than-human performance level should be easier for these kinds of workloads.

Following the leaders too closely seems like a bad move, at least until a profitable business model for an AI model training company is discovered. Mistral’s models are pretty good, right? I mean they don’t have all the scaffolding around them that something like chatGPT does, but building all that scaffolding could be wasted effort until a profitable business model is shown.

Until then, they seem to be able to keep enough talent in the EU to train reasonably good models. The kernel is there, which seems like the attainable goal.

>Mistral’s models are pretty good, right

Are they? IIRC their best model is still worse than the gpt-oss-120B?

Devstral 2 should be above https://mistral.ai/news/devstral-2-vibe-cli

Though I haven't checked other benchmarks and they only report swe

Devstral 2 is free from the API. That has to be a bigger point to what makes it better. The price to performance ratio is practically better in every way. Does it matter if the performance is slightly worse when it is practically free?

> It seems like EU in general should be heavily invested in Mistral's development, but it doesn't seem like they are

The EU is extremely invested in Mistral's development: half of the effort is finding ways to tax them (hello Zucman tax), the other half is wondering how to regulate them (hello AI act)

I guess it's better to do the same stuff everyone else is doing?

I think there is a lot of broad support, but they're just kind of hamstrung by EU regulation on AI development at this stage. I think the end game will ultimately be getting acquired by an American company, and then relocating.

I hope the EU blocks any acquisitions by American companies. The west needs to start protecting its strategic assets.

Do you have any source on this other than vibes based on”EU bad” sentiment?

>It seems like EU in general should be heavily invested

Maybe, i think it will be to our benefit when the bubble pops that we are not heavily invested, no harm investing a little.

> Mistral OCR 3 is ideal for both high-volume enterprise pipelines and interactive document workflows.

I don’t know how they can make this statement with 79% accuracy rate. For any serious use case, this is an unacceptable number.

I work with scientific journals and issues like 2.9+0.5 and 29+0.5 is something we regularly run into that has us never being able to fully trust automated processes and require human verification every step.

Those are tricky! We've found https://www.datalab.to/ to be good for this @ thesynthesis.company

Where are you seeing 79% accuracy? 79% only occurs on the page as a win rate, not an accuracy

And I believe the number is 74%, compared to OCR 2.

What matters is whether this is better than competition/alternatives. Of course nobody is just going to take the output as is. If you do that, that's your problem.

79% win over OCR2 was just for English.

Right! I didn’t know the difference. Does it mean for 79 out of 100 documents they produce 100% accurate OCR, I doubt it. The win rate sounds like a practical approximation of accuracy here to me.

If I am wildly off, I am happy to learn.

79 out of 100 documents Mistral OCR 3 provides better output than Mistral OCR 2.

79% of the time it beats the previous model.

The previous version already achieved up to 99% accuracy in multiple benchmarks, already better than most OCR software.

Thank you.

there has been so many open source OCR in the last 3 months that would be good to compare to those especially when some are not even 1B params and can be run on edge devices.

- paddleOCR-VL

- olmOCR-2

- chandra

- dots.ocr

I kind of miss there is not many leaderboard sections or arena for OCR and CV and providers hosting those. Neglected on both Artificial Analysis and OpenRouter.

Someone posted a project here about a month ago where they compare models in head-to-head matchups similar to llmarena

https://www.ocrarena.ai/leaderboard

Hasn't been updated for Mistral but so far gemeni seems to top the leaderboard.

OCR developers from decades past must be slapping their foreheads now that it seems users will wait a whole minute per page and be happy.

What they are happy about is accurate OCR.

Getting the wrong answer really quickly is not the best goal.

You can also sort by latency. dots.ocr has the lowest at 3.8s/page. And although it doesn't fare very well against much larger slower models, it's still streets ahead of traditional OCR techniques

How can something have a very high ELO but a very low win rate?

You don't loose any elo if your opponent is much stronger than you. Remis could in theory play a part as well.

what I like in MistralOCR is that they have simple pricing $1/1k pages and API hosted on their servers. With other OCR is hard to compare pricing because are token based and you don't know how many tokens is the image unless you run your own test.

E.g. with Gemini 3.0 flash you might seem that model pricing increased only slightly comparing to Gemini 2.5 flash until you test it and will see that what used to be 258 per 384x384 input tokens now is around 3x more.

But they doubled the price g for this new mistralocr3 model to 2$

Simple would be to bill per character.

Now I have to figure out how large a page can be.

I spent like three hours trying to get one of these running and then gave up. I think the paddleOCR one.

It took an hour and a half to install 12 gigabytes of pytorch dependencies that can't even run on my device, and then it told me it had some sort of versioning conflict. (I think I was supposed to use UV, but I had run out of steam by that point.)

Maybe I should have asked Claude to install it for me. I gave Claude root on a $3 VPS, and it seems to enjoy the sysadmin stuff a lot more than I do...

Incidentally I had a similar experience installing open web UI... It installed 12 GB of pytorch crap.. I rage quit and deleted the whole thing, and replicated the functionality I actually needed in 100 lines of HTML.... Too bad I can't do that with OCR ;)

gemini-cli is good for this sort of thing. You can just tell it "Find out why xyz.py doesn't run" and let it crunch. It will try reasonably hard to get you out of Python dependency hell, and (more important) it generally knows when to give up.

But yes, in general, you want to use uv. Otherwise, the next Python application you install WILL break the last one you installed.

I suppose you could use gemini-cli as a substitute for proper Python virtual environment management, always letting it fix whatever broke since the last time you tried to run the program, but that'd be like burning down a rainforest to toast a marshmallow.

https://www.codesota.com/ocr

[dead]

This might be a good place to check the options available for OCR in-place translations. I took a look at OCR3, but it doesn't seem to support my use-case. It looks more tailored towards data extraction for further processing.

I've got some foreign artbooks that I would like to get translated. The translations would need to be in place since the placement of the text relative to the pictures around it is fairly important. I took a look at some paid options online, but they seemed to choke - mostly because of the non-standard text placements and all.

The best solution I could come up with is using Google Lens to overlay a translation while I go through the books, but holding a camera/tablet up to my screen isn't very comfortable. Chrome has Lens built in, but (IIRC) I still need to manually select sections for it to translate - it's not as easy to use as just holding my phone up.

Anyone know of any progress towards in-place OCR/translations?

If you don't mind a paid solution, try DEEPL. I also use Word's built in document translation to good effect.

I don't mind paying for one, though I do remember trying DEEPL without much success. Can't remember the problem offhand, but one of the services I tried just gave me a generic error when I uploaded the PDF. My view at the time was that it had a conniption and just gave up.

Wonder if Word uses the same system Edge has. I remember Edge was also good, but like Chrome's Lens, I'd need to highlight sections for it to get translated. Edge also OCR'd everything very well - just didn't do the translation part automatically.

I am testing it as a replacement of MathPix, first few tests look rather decent. In python for windows: https://pastebin.com/uyiFHKdJ (alpha version prototype). Launches windows snip tool, waits for clipboard image, calls Mistral, retrieves markdown and puts it as text in the clipboard, ready to be pasted in Typora, Obsidian, or other markdown editor.

My main beef with mistral is that they don’t bother to respond to customer inquiries for products the hide behind “reach out for pricing” terms, so even if they were better than SoTA it wouldn’t really matter.

I absolutely loathe dealing with sales people.

I will pay a premium for an inferior product or service if it means I don't have to deal with sales people.

Agreed. In this case the offering just fit neatly into a non core stack we had designed and displaced a bunch of stuff didn’t want to build ourselves.

I also hate dealing with sales people and am not going to reach out to them via another avenue as they will try and posture as if they’re doing us a huge favor (in contrast to me begging gdb for gpt4 api access).

No one mentioning the possibly most beautiful css effect on the Internet??

Is open router still sending all OCR jobs to Mistral? I wonder if they're trying to keep that spot. Seems like Mistral and Google are the best at OCR right now, with Google leading Mistral by a fair bit.

(I work at OpenRouter) If you send a PDF to our API we will:

1. Use native PDF parsing if the model supports it

2. Use this Mistral OCR model (we updated to this version yesterday)

3. UNLESS you override the "engine" param to use an alternate. We support a JS-based (non-LLM) parser as well [0]

So yes, in practice a lot of OCR jobs go to Mistral, but not all of them.

Would love to hear requests for other parsers if folks have them!

[0] https://openrouter.ai/docs/guides/overview/multimodal/pdfs#p...

Hey, I'm the founder of Datalab (we released Chandra OCR). I see someone requested it below - happy to help you all get setup. I'm vik@datalab.to

That links gives an error and so does https://openrouter.ai/docs/guides/overview/multimodal/pdfs

Chandra

Not OS / free weights right?

Gave it a birth registry from a Portuguese locality from 1755 which my dad and I often decipher to figure out geneology and it did a terrible job.

Regular Gemini Thinking can actually get 70-80% of the documents correct except lots of mistakes on given names. Chatgpt maybe understands like 50-60%.

This Mistral model butchered the whole text, literally not a word was usable. To the point I think I'm doing something wrong.

The test document: https://files.fm/u/3hduyg65a5

Just gave it a shot with Grok 4.1 thinking - do you have the ground truth translation to compare? I've tried 4 different times, with slight tweaks adding information from your description, and it's given me a range of interpretations. It'd be nice to see if any of them got close - a couple were more like pulpy telenovela plots, lol.

The model might need tuning in order to be effective - this is normal for releases of image mode models, and after a couple days, there will be properly set up endpoints to test from, so it might be much better than you think. Or it could be really bad with turn of the 19th century portugese cursive.

Oh god, I'm sure I wouldn't come close to 50%; that's so hard to read

It's tough but my dad is quite good at it. He has books of common abbreviations and agglutinations from different centuries. After you get used to it it's faster and very fun.

We were mind blown how good Gemini was at it.

I am too. Gemini 3.0 fast on old scrawled diary entries in English from 100+ years ago got them 95% right. It also added historical context when I prefaced the images with the identity of the writer, such as summaries of an old military unit history in Europe post-WW1 it got from a very obscure U.S. Army archive.

Huge timesaver.

Quick tip: when you digitize a page, put a sheet of black paper behind it. That keeps the ink on the other side from bleeding through.

Forgivable, as that's a quite atypical document, I'd say.

Not atypical enough for Gemini is my point. Also its one of the most common hand written document types in existance since at the time almost nobody other than the local priest knew how to write and birth and marriage certificates were probably the only written documents in whole towns and villages. This is the same throughout Europe at least.

[flagged]

You might want to mention that you are a competitor.

thought you were being flip or assuming that, but checked their profile and you are right. I agree that this should be disclosed in their comment.

How do you know that?

All of their comments are either a) "this OCR is bad", or b) "here is my team's OCR, it is good". Quite blatant.

At instances where data accuracy is of paramount importance, i think a hybrid route of non-llm ocr for data parsing and LLMs for structured data extraction is the safe passage to tread on. Seen better results for LLMWhisperer(OCR)[1] and Latest Gemini.

[1] - https://pg.llmwhisperer.unstract.com/