Skip to content
#19ASep 12, 2026Evidence B

DeepSeek-V4-Flash-Vision-Exp

DeepSeek · 305B total / 13B active (MoE)

An MIT-licensed vision experiment for documents

A vision experiment from DeepSeek released in September 2026: 305B-A13B, 256K, MIT, with documents 94, charts 90, perception 91 and a composite of 87.4. The “Exp” matters — it is an experimental branch with no video or audio, while agents 80 and generation 70 are merely serviceable.

TextImageNative

Last updated Oct 21, 2026Editorial scoring

Bottom line

This is a documents-plus-charts component, not a general multimodal workhorse: documents 94, MIT, ¥0.8/1M output tokens make it strong value for batch extraction. Video, audio, generation and agents are all clearly weak, and as an experimental branch its version stability is your problem.

Eight-dimension breakdown

editorial mapping · 0–100
Perception91Document94Chart QA90Video82Agent80Generation70Chinese88Ecosystem948D
  • DeepSeek-V4-Flash-Vision-Exp87.4
Visual perception & grounding20%
91
Document parsing & Chinese OCR18%
94
Chart & document reasoning14%
90
Video & long-horizon reasoning14%
82
GUI agents & tool use14%
80
Generation & editing6%
70
Chinese context & culture6%
88
Ecosystem, cost & availability8%
94

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
参数规模305B-A13BOfficial spec
许可证MITCommercial use
上下文256KOfficial spec

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 OCR94 · Strong on print and tables
Chart data extraction90 · Reliable on multi-series charts
30-min video QA55 · No video support
Mobile GUI control70 · Not GUI-specific
Chinese cultural context88 · Strong technical context; average everyday context

Strengths and limitations

Strengths

  • Documents 94, MIT and ¥0.8 make batch extraction friendly on both cost and legal terms.
  • Ecosystem 94: DeepSeek’s tooling and hosted API are mature.
  • 256K context keeps long contracts and multi-page reports from over-chunking.

Limitations

  • No video or audio; only text and image are covered.
  • Generation 70 and agents 80: unsuitable for creative work or complex orchestration.
  • An experimental branch with no stability or long-term support commitment.

Price and real cost

List price

¥0.8/M tokens

≈ ¥0.8 / 1M output tokens; weights free. 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-V4-Flash-Vision-Exp

Is DeepSeek-V4-Flash-Vision-Exp open source?

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

How much does DeepSeek-V4-Flash-Vision-Exp cost?

≈ ¥0.8 / 1M output tokens; weights free. It ranks #19 with a composite of 87.4. Real bills depend on resolution, tiling, caching and volume discounts.

What is DeepSeek-V4-Flash-Vision-Exp best for?

Batch contract and report extraction; On-prem document pipelines under MIT; Chart extraction and structuring。

Why does DeepSeek-V4-Flash-Vision-Exp score 87.4?

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.