DeepSeekOpen Source

DeepSeek V4-Flash-Vision-Exp

Compare this model

DeepSeek's first experimental multimodal vision model in the V4 family, open-sourced on Hugging Face under the MIT license on August 31, 2026. It adds a vision encoder and aligner on top of the V4-Flash architecture, gaining image understanding while keeping text-only agent performance on par with V4-Flash. On Terminal-Bench 2.1 it scores 83.9 versus Opus 4.8's 85.0, and leads on DeepSWE (59.3 vs 58.0), signalling near-Opus-4.8 multimodal agent capability.

Parameters

305B

Context Window

TBA

License

MIT

Release Date

2026-08-31

API Pricing

Input Price (per 1M tokens)

$0.14

Output Price (per 1M tokens)

$0.28

Billing Mode: standard

Strengths

  • Keeps text-only agent performance on par with V4-Flash
  • Near Opus 4.8 on Terminal-Bench 2.1 (83.9 vs 85.0)
  • Open weights under MIT license enable local deployment
  • Strong on image understanding, chart analysis, and visual UI agent tasks

Weaknesses

  • Research-stage release not yet deployed by any inference provider
  • Trails Opus 4.8 on harder text tasks like NL2Repo (57.7 vs 69.7)
  • Context window size not disclosed
  • OCR accuracy and complex UI operation still need real-world validation

Use Cases

  • Understanding image-bearing documents, charts, and screenshots
  • Multimodal agent automation that requires visual input
  • Privacy-sensitive inference via local deployment