Slight tangent: i was wondering why DeepSeek would develop something like this. In the linked paper it says

> In production, DeepSeek-OCR can generate training data for LLMs/VLMs at a scale of 200k+ pages per day (a single A100-40G).

That... doesn't sound legal