Mistral co-founder Guillaume Lample demonstrates Mistral OCR 4 converting a handwritten calculus exam into LaTeX

http://atomic.chat

@atomic_chat_hq Handwriting to LaTeX in one shot. My professors couldn't even read my handwriting and they had the answer key.

@tb18_08 because a formula maps nicely to LaTeX, while a graph has to be interpreted first. OCR can read the image, but rebuilding the underlying data and chart structure is a much harder problem

@atomic_chat_hq At least it tagged the chart. Most OCR just ghosts the visuals entirely. Small win.

@atomic_chat_hq Why is it so hard to solve graphs in ocr?

@atomic_chat_hq have you’ll tested the others?

@atomic_chat_hq That's awesome! Tech really is transforming the way we handle exams. Can't wait to see what's next!

@jaseayoub student handwriting recognition is the real AGI test haha

@atomic_chat_hq this is really expensive

@atomic_chat_hq unless they need more than an 8xH100 cluster per OCR model instance (???) that's a diabolic API markup, given the cost/inference duraion ratio

@atomic_chat_hq 5.1s for such accurate LaTeX is impressive

@atomic_chat_hq what the...

@atomic_chat_hq 🔥

@geetkhosla not yet. we only tested the new model in this run

@Dipanshu_AI absolutely! going from handwritten notes to structured digital content in seconds opens up a lot of possibilities for education and research

@atomic_chat_hq That chart tagging is a smart fallback. For the redraw, try piping the LaTeX into a vector graphics tool next. Keeps the math precise while fixing the visual gap. Solid engineering trade-off there.

@atomic_chat_hq Love it!

@atomic_chat_hq does the rate compare with tessaract?

@atomic_chat_hq tagging the chart is huge. structured data beats raw text dump any day for downstream parsing.

I’d nuance that. Both tasks map printed visual information into structured code, so both require interpretation. Formulas are easier because the syntax is mostly symbolic/local. Graphs add axes, scale, legends, encodings and data reconstruction; harder, yes, but not a fundamentally different problem.