Skip to content
#20B+Jan 27, 2026Evidence A

DeepSeek-OCR 2

DeepSeek · Undisclosed (DeepEncoder V2)

A specialist that makes OCR nearly lossless

Scores 91.09 on OmniDocBench v1.5, up 3.73 points, reading documents in human causal order. Not a generalist — but on documents it is in a league of its own.

TextImageBolt-on

Last updated Oct 18, 2026Editorial scoring

Bottom line

Its composite score misleads: narrow modality coverage and no video/agent work cap it at B+. But for turning scans into structured data, it is the first Chinese model to try in 2026 — an order of magnitude cheaper.

Eight-dimension breakdown

editorial mapping · 0–100
Perception84Document99Chart QA88Video70Agent72Generation60Chinese90Ecosystem768D
  • DeepSeek-OCR 281.9
Visual perception & grounding20%
84
Document parsing & Chinese OCR18%
99
Chart & document reasoning14%
88
Video & long-horizon reasoning14%
70
GUI agents & tool use14%
72
Generation & editing6%
60
Chinese context & culture6%
90
Ecosystem, cost & availability8%
76

Weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8. Definitions and sources on the methodology page.

Public benchmarks

as reported, not converted
Benchmark / metricValueNote
OmniDocBench v1.591.09%+3.73 vs previous
单价¥0.8Per 1M output

Public scores come from different versions and prompts and are not interchangeable with our composite; they are one evidence input.

Scenario mapping

five stress tests, one scale
ScenarioScoreObservation
Chinese receipt OCR98 · Best in test
Chart data extraction87 · Great layout parsing, average arithmetic
30-min video QA60 · Unsupported
Mobile GUI control65 · Reads screens, cannot plan
Chinese cultural context88 · Strong document context

Strengths and limitations

Strengths

  • OmniDocBench 91.09 — #1 on documents.
  • Reading order and layout restoration are close to human.
  • Open and very cheap for large batch jobs.

Limitations

  • Image/document only; no video or audio.
  • Generation and agent abilities are largely absent.
  • Handwriting and low-quality scans still need work.

Price and real cost

List price

¥0.8/M tokens

Open; API ≈ ¥0.8 / 1M output. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥0.5 (assuming input at 25% of output price).

Use your volume

About DeepSeek-OCR 2

Is DeepSeek-OCR 2 open source?

Licence: Open (MIT-style). Weights are available, but check the exact terms and filing requirements before commercial use.

How much does DeepSeek-OCR 2 cost?

Open; API ≈ ¥0.8 / 1M output. It ranks #20 with a composite of 81.9. Real bills depend on resolution, tiling, caching and volume discounts.

What is DeepSeek-OCR 2 best for?

Batch receipt and contract structuring; Archive digitization; RAG document preprocessing。

Why does DeepSeek-OCR 2 score 81.9?

The composite weights eight dimensions: perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. Its strongest dimension is Document parsing & Chinese OCR; its weakest is Generation & editing.

Personal ranking · Not an official score

Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.