← All models

DeepSeek-OCR 2

Published Publisher DeepSeek

Overview

DeepSeek-OCR 2 is an open-weight vision-language model for optical character recognition and document understanding. DeepSeek announced it on January 27, 2026. It introduces DeepEncoder V2, which can reorder visual information according to document semantics rather than relying solely on a fixed raster-scan order. (github.com)

Capabilities and uses

  • Recognizes text from document images and can produce Markdown with layout grounding; it also supports OCR without layout markup.
  • Official inference examples cover images, concurrent PDF processing and batch evaluation. Potential uses include document digitization and preparing document data for language-model pipelines.
  • The official repository provides vLLM and Transformers inference examples. Code and model weights are available under Apache 2.0. (github.com)

Limitations

  • Recognition, layout and reading order can contain errors; outputs should be checked when accuracy matters. DeepSeek's paper reports that repetition errors remain, although they decreased compared with the original model in its evaluated production workloads.
  • General visual reasoning and more advanced two-dimensional reasoning are research directions, not guaranteed capabilities. The supplied local inference examples use CUDA and specific software dependencies. (arxiv.org)

Pricing

Model-specific API pricing: Not verified in DeepSeek's official Chinese or international pricing documentation as of November 17, 2026. Separate input-token and output-token prices for DeepSeek-OCR 2 have not been verified. (api-docs.deepseek.com)

Related service, not a verified model-specific quote: DeepSeek's official OCR API is priced by successfully recognized pages, not separately by input and output tokens. Its Chinese documentation lists CNY 1.5 per 1,000 pages; its international documentation lists USD 0.225 per 1,000 pages. Failed or timed-out recognition is not billed. The documentation does not establish that this service uses DeepSeek-OCR 2, so these are service prices, not confirmed prices for this model. Checked November 17, 2026. (api-docs.deepseek.com)

Official sources