OCR Arena is a free playground for evaluating leading VLMs and OCR models side-by-side. Upload any document, compare accuracy, and vote for the best models on a public leaderboard.
Hey Product Hunt 👋!
Kushal here, CEO of Extend. At Extend, we're building the world's best document processing platform. Today, we're excited to share OCR Arena with the community.
OCR Arena is a free playground for evaluating open source OCR models and foundation VLMs side-by-side. Upload any document, compare accuracy, and vote for the best models on a public leaderboard.
OCR is going through a golden era. Every week, it feels like a new open source model comes out and sets a new record. But testing them is still painful.
Academic benchmarks only tell part of the story, but ultimately teams care most about how models perform on their specific documents and edge cases. Our goal with OCR Arena is to reduce the friction of testing new models and make OCR evaluation open, unbiased, and grounded in real-world performance.
We’ve initially launched with 10+ models, from Gemini 3 to DeepSeek-OCR to Qwen3-VL (powered by our friends over at Baseten!). If there's any model missing that you'd like to see, let us know in the comments and we'll do our best to get it live quickly. And as new models are released, we'll add them to the arena so they can compete for a spot on the leaderboard.
Grab a messy PDF, head over to OCR Arena, and see which models work best!
We'd love to hear any and all feedback on how we can make this better for the community. We'll be here all day to answer your questions.
Thanks for checking us out!
Feature request: it would be nice if there is a filter which only shows open source models, a filter to pick models greater than a size. This would ease the process for people working with open-source models.
Useful tool for teams comparing models. I’m curious how well OCR Arena handles very noisy or low-resolution documents, does the leaderboard shift much when you test harsher real-world cases?
Very fun little arena! I have been using VLMs for OCR tasks in production, so I am drawn to it and want to see which model excels at the task of receipt scanning, which is the main feature of an app @ReceiptGenie (launching on Nov 23) I have been working on.
It is surprising to find that the model I am using is ranking quite low in the Arena, and my real results (after almost 5k scans) were basically on par with the best performant model in the Arena. I feel like a solid prompt can go a long way for specific tasks.
I am very surprised by the performance of the Qwen3-VL-8B model. Definitely packs a good punch for the size. Will definitely experiment a bit more on this and see how much I can steer the model with prompting.
Would be good if we can see a few more OSS models like MinerU etc, and being able to test the models side by side with the same prompt would be very cool.
Hi everyone 🙂
I'm Ishaan, the dev who built this!
We put a lot of time and care into building OCR Arena and I really hope you all love it. Please let me know if you have any feature requests, new models you want to see, or anything else you think that might make this a fun and useful product to use!
