DeepSeek-OCR compresses text as images using up to 10x fewer tokens

DeepSeek today unveiled DeepSeek-OCR, a model that converts large amounts of text into compact visual representations before recovering it through optical character recognition. The experiment achieves 97% accuracy while using up to roughly ten times fewer vision units than text tokens—a potential way to cut the cost of processing long documents and rethink how language models store context.

The work, published as an arXiv preprint along with the model’s code and weights, does not show that images will replace text as the standard input for assistants. It does provide an important proof of concept: a page can carry a great deal of information through a relatively small number of vision tokens, the units a model uses to divide and process an image.

Turning a long sequence into a two-dimensional map

Language models do not read words the way people do. They first break text into tokens, which may correspond to a word, part of a word or a punctuation mark. The longer the document, the more tokens must be processed and kept in the context window, increasing memory and computing costs.

DeepSeek-OCR takes a different approach. It renders the content as a two-dimensional image and uses a component called DeepEncoder to condense it into vision tokens. A decoder called DeepSeek3B-MoE-A570M then reconstructs the content using OCR, the technology used to extract text from images and scanned documents.

The decoder uses a mixture-of-experts, or MoE, architecture: it has different specialized blocks but activates only some of them for each input. Its name indicates roughly 3 billion total parameters, with around 570 million active during processing.

The appeal of visual representation lies in its density. An image can combine text, columns, tables, formulas and spatial relationships in a grid. Converted into a linear sequence, describing that structure requires many tokens or additional labels.

97% accuracy at nearly 10x compression

In the team’s experiments, DeepSeek-OCR retains 97% decoding accuracy when the original number of text tokens is no more than ten times the number of vision tokens. In other words, a representation of roughly 100 vision tokens could condense content that would take up nearly 1,000 text tokens, although the exact ratio depends on the document and how it is rendered.

Accuracy declines as compression becomes more aggressive. At a 20x ratio, it drops to around 60%. That decline marks a practical limit: an image is not an unlimited container, and details eventually disappear when its information density is reduced too far.

DeepSeek also compares the model on OmniDocBench, a benchmark for document-understanding systems. DeepSeek-OCR outperforms GOT-OCR 2.0 using 100 vision tokens per page, compared with 256 for the competing system. It also delivers better results than MinerU 2.0 with fewer than 800 tokens, while the latter uses more than 6,000 per page on average, according to the paper.

More than 200,000 pages a day on a single A100

The team says the system can process more than 200,000 pages a day with a 40GB Nvidia A100 GPU. That works out to slightly more than two pages per second if the machine runs continuously, although the figure alone does not represent the full cost of a production infrastructure.

The immediate use case is preparing data to train language and multimodal models. Companies digitizing archives, invoices, reports or scientific literature could process documents with fewer vision tokens and generate large training corpora using a single data-center GPU.

It could also be useful for pages whose meaning depends on layout. A text-only system generally needs to convert tables, charts and positions into an intermediate format. Visual input preserves that structure from the outset, although it requires the model to correctly interpret small type, different resolutions and unusual layouts.

Fewer tokens does not automatically mean 10x lower costs

Comparing vision tokens with text tokens requires caution. The two do not necessarily have the same computational cost: before reaching the decoder, the image must be rendered and passed through the vision encoder. As a result, a 10x reduction in token count does not by itself imply a 10x saving in time, memory or money.

The reported 97% also measures the ability to reconstruct text under specific experimental conditions. It does not show that a model reasons equally well about that compressed content, remembers all its details during a long conversation or maintains its accuracy across every language and document format.

DeepSeek proposes a possible future direction for model memory: storing older information at lower resolution while reserving more capacity for recent context. The analogy to human forgetting is appealing, but it remains a research hypothesis. The next step will be to determine whether optical compression preserves not just characters, but also the specific data a model needs to answer questions, compare information and reason about long documents.