AI Radar

DeepSeek-V4-Flash-Vision-Exp

Available · Language, Multimodal, Reasoning

DeepSeek-V4-Flash-Vision-Exp is an experimental model that DeepSeek calls its first multimodal model in the DeepSeek-V4 family, built on the V4-Flash architecture with added visual modules. It accepted text and images in the DeepSeek API from 2026-08-21, later got open weights under the MIT License and was retired in the API on 2026-09-10. [1]Introduction; License [2] [3]Date: 2026-09-10 Primary source

Timeline of DeepSeek-V4-Flash-Vision-Exp →

Claims and evidence

  • Released API release; the weights appeared on Hugging Face later (repository created 2026-08-31 according to hub metadata; no page states the date). Primary source[2] [3]Date: 2026-08-21
  • Status AvailableWeights downloadable from the deepseek-ai Hugging Face organisation. In the DeepSeek API the model was retired on 2026-09-10; the legacy name deepseek-v4-flash-vision-exp now routes to DeepSeek-V4.1-Flash. Primary source[1]License [3]Date: 2026-09-10
  • Replaced by DeepSeek-V4.1-FlashIn the DeepSeek API: retired on 2026-09-10; deepseek-v4-flash-vision-exp temporarily routes to DeepSeek-V4.1-Flash. Primary source[3]Date: 2026-09-10 [4] [5]
  • Derived from (other) DeepSeek-V4-FlashThe card says it builds on the DeepSeek-V4-Flash architecture by adding visual modules and continued training. It does not name the starting checkpoint; the card compares it with DeepSeek-V4-Flash-0731 and the change log with the official DeepSeek-V4-Flash, so that revision is a possible alternative target. Primary source[1]Introduction
  • Change · Modality Adds image input next to text; images can be sent as base64, as external URLs or through the new Files API.Compared with DeepSeek-V4-Flash Primary source[2]Multimodal API support
  • Change · Architecture Adds visual modules to the V4-Flash architecture, followed by continued training.Compared with DeepSeek-V4-Flash Primary source[1]Introduction
  • Change · Availability Offered only in the DeepSeek API at launch, as an experimental model; the weights followed later on Hugging Face.Compared with DeepSeek-V4-Flash Primary source[2] [1]License
  • Input text, image Primary source[2]Multimodal API support [1]Introduction
  • Output text Primary source[1]pipeline tag image-text-to-text
  • Feature Reasoning modeThe card reports evaluating it at the max reasoning effort level. Primary source[1]Introduction, note 1
  • Open weights YesRepository licensed under the MIT License; weight shards published on Hugging Face after the API release. Primary source[1]License; Repository layout
  • Access API, Open-weights download Primary source[2] [1]License

Lineage

Predecessors

No known predecessor.

Successors

No known successor.

Based on

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from DeepSeek on AI Radar →