We outperformed Mistral OCR with a tiny team | Hacker News Reader

We outperformed Mistral OCR with a tiny team

7 pointsapichar6 comments

Mistral recently launched their OCR model, claiming to be the best in the world. We put that claim to the test against JigsawStack vOCR — our own OCR solution, built by a team of just three people.

We benchmarked both models across real-world multilingual documents, handwriting recognition, structured data extraction, and bounding box accuracy. Here’s a quick breakdown:

- Languages: Mistral supports 12 benchmarked languages, while JigsawStack vOCR handles 70+ including Telugu, Hindi, and lesser-used scripts.

- Handwriting Recognition: Mistral struggles with handwritten and distorted text, while JigsawStack vOCR accurately extracts text from printed materials, handwriting, and even text on walls.

- Structured Output: Mistral requires additional post-processing with an LLM for structured data, while JigsawStack natively returns structured JSON output.

- Bounding Boxes: Mistral OCR does not provide bounding box data, while JigsawStack supports both sentence and word-level positions.

Check out the full breakdown with examples, screenshots, and API comparisons in our blog post Mistral OCR vs. JigsawStack vOCR here: https://jigsawstack.com/blog/mistral-ocr-vs-jigsawstack-vocr

We'd love to hear feedback and answer any questions! If you’re building with OCR, try out JigsawStack vOCR and let us know your thoughts.

Your colleague ykjs already told us: https://news.ycombinator.com/item?id=43289571

True- I saw that too, but I also wanted to start a thread to chat with the community and hear different perspectives. Would love to get your thoughts!

is yours more expensive? on your pricing page, your plan is 0.05 per invocation. how many pages can be handled by a single invocation?

Our pricing model is being updated soon to be even more cost-effective. Currently, we support up to 10 pages per API request. With the upcoming token-based pricing, costs will drop significantly to just $1.40 per 1M tokens.

how do a client break a pdf into 10 page chunks? can a pdf file be uploaded, or we are expected to upload rasterized images?

I also feel that token is not a comprehensible unit for many customers. "page" is better.

You can upload a PDF directly, and the vOCR API supports a page_range parameter so you can specify which pages to process in a single request (e.g., [1,10]).

If your PDF has more than 10 pages, you can make multiple requests, each specifying a different page range (e.g., [1,10], [11,20], etc.). No need to manually split the file.

Also, I totally get that tokens can feel abstract. However by making this shift we can focus on processing time. In doing so the token-based pricing model allows us to be more flexible and provide more at a cheaper cost.

For example, instead of $0.05 per invocation it could be as low as $0.0014–$0.0035 per page depending on factors like the amount of text extracted and formatting.

We do appreciate the feedback and will continue make this approach as transparent as possible.

Reply on news.ycombinator.com