Unlimited OCR: One-shot long-horizon parsing

	Unlimited OCR: One-shot long-horizon parsing(github.com)
	475 points by ingve 1 day ago \| 108 comments
	tl;dr: Baidu released Unlimited-OCR, a document parsing model that extends DeepSeek-OCR for one-shot long-horizon parsing of single images, multi-page documents, and PDFs up to 32k tokens. It supports two image configurations (gundam and base) and ships with both Hugging Face Transformers and SGLang inference paths, including a batch script with concurrent requests against an OpenAI-compatible API. The model is available on Hugging Face and ModelScope, with an accompanying arXiv paper.
	HN Discussion: ↑Technical appreciation for the architectural approach to solving KV cache memory issues in long document OCR ↑Enthusiasm for multi-page single-pass VLM OCR with interest in the attention mechanism design •Skepticism about OCR hallucinations and whether this model avoids inventing artifacts •Questioning how this compares to other OCR tools like Infinity Parser 2 or Mistral's offering •Curiosity about why companies open-source genuinely valuable software like this