According to Andrej Karpathy (@karpathy), the new DeepSeek-OCR paper presents a notable advancement in OCR models, though slightly behind state-of-the-art models like Dots. The most significant insight lies in its proposal to use pixel-based image inputs for large language models (LLMs) instead of traditional text tokens. Karpathy emphasizes that image-based inputs could enable more efficient information compression, resulting in shorter context windows and higher computational efficiency (source: Karpathy on Twitter). This method also allows LLMs to process a broader range of content—such as bold or colored text and arbitrary images—with bidirectional attention, unlike the limitations of autoregressive text tokenization. Removing tokenizers reduces security risks and avoids the complexity of Unicode and byte encoding, streamlining the LLM pipeline. This vision-oriented approach could open up new business opportunities in developing end-to-end multimodal AI systems and create more generalizable AI models for enterprise document processing, security, and accessibility applications (source: DeepSeek-OCR paper, Karpathy on Twitter).